IMPROVING STEAM TEMPERATURE CONTROL WITH NEURAL NETWORKS
by
JACQUES FRANCOIS SMUTS
THESIS
submitted in partial fulfilment of the degree
DOCTOR INGENERIAE in MECHANICAL ENGINEERING at the RAND AFRIKAANS UNIVERSITY
SUPERVISOR: Prof. A.L. NEL
JULY 1997 Summary The thesis describes the development, installation, and testing of a neural network-based steam • temperature controller for power plant boilers. Attention is focussed on the mechanical and thermodynamic aspects of the control problem, on the modelling and control aspects of the neural network solution, and on the practical and operational aspects of its implementation. A balance between theoretical and practical considerations is strived for. Experimental data is obtained from an operational coal fired power plant.
As a starting point, the importance of good steam temperature control is motivated. The sensitivity of heated elements in boilers to changes in heat distribution is emphasized, and it is shown how various factors influence the heat distribution. The difficulties associated with steam temperature control are discussed, and an overview of developments in advanced steam temperature control on power plant boilers is given.
The suitability of neural networks for process modelling and control are explored and the error backpropagation technique is shown to be well suited to the steam temperature control problem. A series of live plant tests to obtain modelling data is described and specific attention is given to discrepancies in the results. The prOcess of selecting the ideal network topology is covered and improvements in modelling accuracy by selecting different model output schemes are shown.
The requirements for improving steam temperature control are listed and the philosophy of optimal heat distribution (OHD) control is introduced. Error backpropagation through the heat transfer model is utilized in an optimizer to calculate control actions to various fire-side elements. The scheme is implemented on a power boiler.
It is shown that the optimizer manipulates control elements as expected. Problems with fuel-to- pressure oscillations and erroneous fuel flow measurement are discussed. Due to process oscillations caused by OHD control, a reduction in control quality is evident during mill trips and capability load runbacks. Substantial improvements over normal PID control however, are evident during load ramps. ii
Opsomming
Hierdie proefskrif beskryf die ontwikkelling, installasie, en toetsing van n neurale netwerk gebaseerde stoomtemperatuurbeheerder vir kragstasieketels. Aandag word gefokus op die meganiese en termodinamiese aspekte van die beheerprobleem, op die modellerings- en beheeraspekte van die neurale netwerk oplossing, en op praktiese- en bedryfsaspekte van die implementering. Daar word gepoog om 'n balans te handhaaf tussen teoretiese en praktiese oorwegings. Eksperimentele data word verkry vanaf 'n operasionele steenkool kragstasie.
As beginpunt word die belangrikheid van goeie stoomtemperatuurbeheer gemotiveer. Verhitte elemente in stoomketels se sensitiwiteit vir veranderings in hitteoordragspatrone word beklemtoon, en daar word aangetoon hoe verskeie faktore die hittebalans beinvloed. Die moeilikhede wat gepaard gaan met stoomtemperatuurbeheer word bespreek, en 'n oorsig van ontwikkelinge in gevorderde stoomtemperatuurbeheer op kragstasieketels word gegee.
Die toepaslikheid van neurale netwerke op prosesmodellering en -beheer word ondersoek en daar word getoon dat die tegniek van fout-terugpropagering gepas is vir stoomtemperatuurbeheer. 'n Reeks toetse wat gedoen is om modelleringsdata te bekom word beskryf, en aandag word spesifiek aan teenstrydighede in die resultate geskenk. Die keuse van 'n ideale netwerkuitleg word gedek en verbeteringe in die akuraatheid van modellering deur middel van verskillende uitsetskemas word getoon.
Die vereistes vir die verbetering van stoomtemperatuurbeheer word genoem en die filosofie van optimale hitteverspreidingsbeheer (OHV beheer) word bekendgestel. Fout-terugpropagering deur die hitteoordragsmodel word gebruik in 'n optimiseerder om beheeraksies aan die vuur-kant te bereken. Die OHV algoritme word op 'n kragstasiestoomketel geimplementeer.
Daar word aangedui dat die optimiseerder die beheerelemente na verwagting verstel. Probleme met brandstof-teenoor-druk ossillasies en foutiewe brandstofmeting word bespreek. As gevolg van prosesossillasies wat veroorsaak word deur OHV beheer, vind 'n daling in beheerkwaliteit plaas gedurende meulklinke en noodgedwonge vragvennindering. Noemenswaardige verbetering bo PID beheer is egter merkbaar gedurende vragveranderinge. iii
Table of Contents
Summary
Opsomming ii
Table of Contents iii
List of Figures vi
List of Tables
List of Variables xi
1. Introduction 1 1.1 Power generation 1 1.2 A brief history of boiler control 2 1.3 The need for steam temperature regulation 5 1.4 Research hypothesis 6 1.5 Overview of thesis 6
The power plant boiler 9 2.1 Cycle description 9 2.2 Heat transfer theory 14 2.3 Steam generator design 19
Steam temperature control 30 3.1 Control elements for steam temperature regulation 30 3.2 Difficulties associated with steam temperature regulation 40 3.3 Temperature excursion study 47 3.4 Instrumentation and control configuration 55 iv
3.5 Developments in steam temperature control 61
4. Neural networks and process control 74 4.1 Description of a neural network 74 4.2 Selecting the size of a neural network 77 4.3 Training the network 78 4.4 Process modelling with neural networks 79 4.5 Process control with neural networks 81
5. Plant modelling 87 5.1 Desired model characteristics 87 5.2 Acquiring test data 89 5.3 Calculations and assumptions 98 5.4 Neural network model 120
Neural networks and steam temperature control 135 6.1 Requirements for improved steam temperature control 135 6.2 Optimal heat distribution control 139 6.3 Controller design 141 6.4 Expected results 155
Practical implementation and results 157 7.1 The PC as control platform 157 7.2 Interfacing to existing boiler controls 159 7.3 Steady state testing and optimization 165 7.4 Transient testing and optimization 167 7.5 Final results 185
Conclusion 190 8.1 Discussion 190 8.2 Return to research hypothesis 192 V
8.3 Future research 193
Bibliography 195
Appendix A. Heat distribution test programme 204
Appendix B. Variables recorded during heat distribution tests 210
Appendix C. Spreadsheet model 213
Appendix D. OHD graphic display 215
Appendix E. Selected test results 216 vi
List of Figures
1.1 South African power demand through a typical day
2.1 Carnot cycle. 9
2.2 Carnot cycle T-S diagram. 9
2.3 Rankine cycle. 10
2.4 Rankine cycle T-S diagram. 11 2.5 Superheat cycle T-S diagram 12
2.6 Reheat cycle with economizer. 13
2.7 Reheat cycle with economizer T-S diagram 13
2.8 Fire-side components of a steam generator. 19
2.9 Different firing systems indicating fuel injection angle: 20
2.10 Diagrammatic view of the water & steam path through power plant components. . 22
2.11 Typical steam temperature characteristics. 23
2.12 Heat rise in boiler elements vs. steam pressure 23
2.13 Different heat zones in a steam generator. 24
2.14 Typical location of steam generator elements 27
2.15 Layout of the Kendal boiler heat transfer elements 29 3.1 The effect of burner tilt angle on fireball elevation. 32
3.2 Effect of burner tilt angle on furnace exit temperatures. 33
3.3 Kendal superheater stages and desuperheater locations. 39
3.4 Reheater outlet temperature reacting to increased spray water flow. 42
3.5 Reheater outlet temperature response under two load conditions 44
3.6 Causes of temperature excursions at Kendal. 48
3.7 Main steam temperature deviations from setpoint caused by load variations 49
3.8 Temperature excursion caused by a mill shut down. 52
3.9 Basic temperature control loop. 55
3.10 Cascade control arrangement. 56
3.11 Feedforward control. 57 3.12 Combined feedback, feedforward and cascade control arrangement. 58
3.13 Multiple control elements with coupled control. 60
4.1 Schematic representation of a typical artificial neuron. 74 vii
4.2 Neuron transfer functions. 75 4.3 Feedforward_ neural network. 76 4.4 Backpropagation signal flow. 84 4.5 Feedforward and backpropagation modes. 85 5.1 Measurements on feed water system, economizer and evaporator. • 92 5.2 Measurements on superheater and reheater. 93 5.3 Correlation between fuel flow and total heat gain was obtained for all tests. 97 5.4 Heat shifts achieved during heat distribution tests. 98 5.5 Feed water heater. 103 5.6 Relation in pressure differential (DP) across superheater stages. 106 5.7 Variables for heat balance calculations 108 5.8 Superheater spray water flow measurement 111 5.9 Reheater spray water flow measurement. 111 5.10 Superheater spray and warmup flow. 111 5.11 Discrepancies between calculated and measured air flow ratio. 115 5.12 Correlation between fuel flow and generator load. 116 5.13 Air flow vs 02 in flue gas with fuel flow derived from generator load. 118 5.14 Normalized difference between LH and RH air flow measurements. 119 5.15 Furnace to boiler heat transfer mapping 120 5.16 Error on test data increases after many training runs. 122 5.17 7:50:3 neural network model output errors for all tests. 127 5.18 Absolute heat transfer rate model. 132 5.19 Relative heat transfer rate model errors. 132 5.20 Corrected relative heat transfer rate model errors. 132 6.1 Model-based predictive control. 136 6.2 Adaptive adjustment concept 137 6.3 Design heat transfer rates to maintain steam temperatures. 138 6.4 Signal flow to and from the optimal heat distribution controller. 140 6.5 Predictive calculation for error in heat manger. 143 6.6 Backpropagation of errors to obtain derivatives. 144 6.7 Bias development during an optimization run. 146 viii
6.8 Heat transfer errors during an optimization run. 146 6.9 Adjusting the design heat transfer to match plant conditions. 152 6.10 Adjusting the heat transfer model to match plant conditions. 154 7.1 Interface between PC and existing boiler control system. 159 7.2 Closed loop control signal flow diagram. 161 7.3 Feedforward control signal flow diagram 162 7.4 Mill bypass damper and air flow paths. 165 7.5 Error estimation on mill fuel flow. 166 7.6 Fuel and steam flow rates during a down ramp under OHD control. 168 7.7 Burner tilt angle during load ramp, showing optimization glitch 169 7.8 Mill demands during ramp, showing biassing error. 169 7.9 Different polynomials fitted to the same three points. 170 7.10 Modelling errors with the 7:15:3 network. 171 7.11 Modelling errors with the 7:5:3 network. 171 7.12 Oscillating fuel flow during down ramp under OHD control. 172 7.13 Burner tilt action to regulate heat transfer to superheater and reheater 172 7.14 Mill biassing to regulate heat distribution 173 7.15 Boiler pressure response to fuel flow with OHD control on and off. 174 7.16 Main steam temperature decreasing during load ramp under OHD control 175 7.17 Predicted and target heat transfer rates to superheater during load ramp. 176 7.18 Discharged and absorbed heat flows. 177 7.19 Mill fuel flow response to increased air through-flow. . 178 7.20 Mill fuel flow response to increased coal input. 178 7.21 Mill fuel flow response to increased coal and air flow. 179 7.22 Fuel flow indication increasing after mill trip. 180 7.23 Correction circuit for mill fuel flow. 180 7.24 Heat discharge calculated from the adjusted fuel flow measurement 181 7.25 Air flow and 02 control. 182 7.26 Deviations in 02 measurement caused by incorrect fuel flow measurement 184 7.27 Effect on 02 on predicted heat discharge. 184 7.28 Reheat spray flow rate used by OHD control to absorb the excess heat transfer. 185 ix
7.29 Biassed mill fuel flows under OHD control compared to normal. 185 7.30 OHD tilt biassing during load ramp. 186 7.31 Heat transfer rate to superheater during down-ramp. 187 7.32 Effect of OHD control on main steam temperature. 187 x
List of Tables
3.1 Mill combinations and corresponding tilt angles 37 3.2 Results of excursion study 47 3.3 Steady state conditions before the ramp 50 3.4 Conditions during ramp. 50 3.5 Changes in heat transfer during load ramp. 51 3.6 Conditions before mill trip 53 3.7 Conditions after mill trip 54 3.8 Changes in heat transfer caused by a mill trip 54. 5.1 Elimination of mill combinations. 90 5.2 Tilt performance: setpoint = -28°, average angle = -21° 99 5.3 Tilt performance: setpoint = 0°, average angle = 1° 99 5.4 Tilt performance: setpoint = 30°, average angle = 20° 99 5.5 Superheater spray water enthalpy. 101 5.6 Turbine outlet and feed water heater inlet conditions 104 5.7 Distillate conditions 105 5.8 Feed water discharge conditions. 105 5.9 Extremes in conditions at first stage desuperheater inlet. 107 5.10 Results of networks trained with 10 individual outputs 125 5.11 Results of networks trained with 3 grouped outputs 126 5.12 Comparison of individual to grouped output heat transfer model. 127 5.13 Comparison of two output strategies. 128 5.14 Improvement in results by modelling relative heat transfer. 129 5.15 Improvement of accuracy by correcting the outputs 130 5.16 Comparison of different heat transfer model results 131 5.17 Summary of results obtained from different network sizes 131 5.18 Heat transfer rates obtained with different initializations. 133 6.1 Improvements in heat transfer after a mill trip. 155 6.2 Furnace element setup after a mill trip. 156 7.1 Accuracy of networks with various numbers of hidden neurons. 171 xi
List of Variables
a boiler tube spacing geometric relation ratio A surface area of a boiler tube
A A Actual air flow rate [kg/s] A s Stoichiometric air flow rate [kg/s]
cpg specific heat of flue gas at constant pressure [J/kg°C] cp ,,,„„, Specific heat of steam at 4 MPa & 420°C [J/kg°C] co meta,Specific heat of 1.5 % carbon steel [J/kg°C]
COIF Concentration of 0 2 in flue gas [%] D dimension of boiler tube surface parallel to gas flow [m] es heat discharge error to evaporator [W] e, heat discharge error to superheater [W] e, heat discharge error to reheater [W] f nonlinear mapping function
fe neural network mapping of evaporator neural network mapping of superheater neural network mapping of reheater
fed design heat transfer curve of evaporator
fsd design heat transfer curve of superheater
frd design heat transfer curve of reheater ha extraction steam enthalpy [J/kg]
/ad distillate (condensed extracted steam) enthalpy [J/kg] lift feed water enthalpy at heater inlet [J/kg] feed water enthalpy at heater outlet [J/kg] h, steam enthalpy at desuperheater inlet [J/kg] ho steam enthalpy at desuperheater outlet [J/kg] h,4„ outlet enthalpy of reheat steam [J/kg] hsp, spray water enthalpy [J/kg] kaa, convection heat transfer coefficient [W/m2 °C] Ica thermal conductivity of ash [W/m°C] kg thermal conductivity of flue gas [W/m°C] k thermal conductivity of a boiler tube [W/m°C] xii
mez extraction steam mass flow rate [kg/s] mf feed water mass flow rate [kg/s] m, steam mass flow rate at desuperheater inlet [kg/s] ink steam leakage rate [kg/s] moo main steam flow rate [kg/s] mo steam mass flow rate at desuperheater outlet [kg/s] desuperheater spray water flow rate [kg/s] Mass of reheater tubing and header material [kg] Mnewo Mass of steam inside reheater tubing & headers [kg] L length of a boiler tube [m] Q quantity of heat [J] P, output of a neuron output of a neuron q heat transfer rate [NV] 9er excess heat transfer [W]
qf total furnace heat discharge [NV] qn actual heat discharge to evaporator qaa actual heat discharge to superheater qn actual heat discharge to reheater [NV] qed design heat discharge to evaporator [W] qsd design heat discharge to superheater [W] q,d design heat discharge to reheater [w] qn predicted heat discharge to evaporator qn predicted heat discharge to superheater [W] qrp predicted heat discharge to reheater [NV] T., heat transfer rate through conduction gram, heat transfer rate through convection [W] grad heat transfer rate through radiation [NW] vector of actual heat transfer rates [W] g a vector of modelled heat transfer rates [W] a,.a vector of corrected modelled heat transfer rates ra outer radius of ash layer [m] r, inner radius of boiler tube [m] output of a neuron r„, scalar sum of relative heat transfer rates ro outer radius of boiler tube [m] vector of modelled relative heat transfer ratios [W/W] 12c film conductance [why? °C] s entropy T temperature [°C] TJ temperature of fluid inside boiler tube [°C] Tg temperature of combustion gas [°C] T,,,„„, Average steam temperature (assumed) [°C] T surface temperature of a boiler tube [°C] Too temperature of a free gas stream [°C] vector of furnace conditions affecting heat transfer rate Vg linear velocity of gas stream [m/s] tr weight (gain) of a neural network connection weight (gain) of a neural network connection w weight (gain) of a neural network connection
WT work produced by turbine [J] We work consumed by compressor [J] input to a neuron or neural network y output of a neuron ae gain factor on the evaporator heat transfer error gain factor on the superheater heat transfer error ar gain factor on the reheater heat transfer error
Pg density of flue gas [kg/m3] boiler thermal efficiency
Pg viscosity of flue gas [kg/ms] a Stefan-Boltzmann constant = 5.669 x 104 [whn2K4] Emissivity of a non black body
1
1. Introduction
1.1 Power generation The world today consumes vast amounts of energy as nations strive to satisfy much more than only the basic human needs of food, shelter and clothing. Virtually the entire environment of a westerner is in some way dependent on adequate supplies of energy. Over the period from 1950 to 1990, annual world electrical power production and consumption rose from slightly less than one trillion kilowatt hours (1.0 * 10' 2 kWh) to more than 11.5 trillion kWh [1].
In South Africa, access to electricity is considered one of the rights of every resident. Eskom, the national power company, with an installed capacity of 38 497 MW, expands its services to new customers at a rate of 300 000 connections per year [2]. This contributed to an average growth in electricity sales of 3.6% over the past five years [2], but also contributed to a growth in the peak electricity demand, with a new winter maximum demand of 27 967 MW recorded on 24 August 1996 [3]. The average power demand during a 24 hour period in South Africa is shown in Figure 1.1. During an average day in the winter, the peak load demand is 50% higher than the base load demand. This demand variation requires many of the power stations to perform large load changes daily.
26
' 24
2 22 E a) 1,3_ 20
pi 18
16 0 3 6 9 12 15 18 21 24 Hour of the day — Summer Winter
Figure 1.1 South African power demand through a typical day. [2] 2
In 1950 rougly two-thirds of the electricity came from thermal (steam-generating) sources and about one-third from hydroelectric sources. In 1990 thermal sources still produced about two-thirds of the power, but hydro power had declined to just under 20 per cent and nuclear energy accounted for about 15 per cent of the total [1]. Of all the fossil fuels used for steam generation in power plants today, coal accounts for most of the energy [4]. At an annual production rate of about 3.5 billion metric tons worldwide, serious depletion of coal resources will take around 185 years [5]. Therefore, it may be said that coal-fired power stations will be one of the prime sources of electrical power for many years to come.
Compared to its beginning, the generation of electricity has become a very complicated business.
High energy costs demand that as much electricity as possible be generated from the fuel . consumed. Higher availability of equipment is needed to stem rising operating and maintenance costs. Protection of both personnel and equipment must be achieved, and unscheduled shutdowns must be kept to a minimum. While obviously instrumentation and control systems cannot satisfy such concerns by themselves, the above demands have resulted in a substantially increased requirement for sophisticated instrumentation and automatic control systems. In this context, modern power plants are among the most highly automated and centrally controlled and monitored production facilities in the world.
1.2 A brief history of boiler control The earliest known boiler control application was that of a float valve regulator for boiler water level control [6]. This device was described in a British patent by James Brindley in 1758. Mother float valve regulator of considerable originality was independently invented in 1765 in Russia by Ivan Polzunov. In a British patent of 1784 Sutton Wood documented some improvements to the float valve regulator. James Watt and Matthew Boulton of Boulton & Watt Co. adopted the float valve regulator as a standard attachment to their boilers somewhere between 1784 and 1791 [6].
A discussion on control system development will probably not be complete without reference to the steam engine governor. The origins of this device lie in the lift-tenter mechanism which was used to control the gap between the grinding-stones in both wind and water mills. Boulton. 3 described the lift-tenter in a letter (dated May 28, 1788) to Watt, who realized it could be adapted to govern the speed of the rotary steam engine. The first design was produced in November 1788, and a governor was first used early in 1789 [7].
Steam pressure control was first patented in 1799 by Matthew Murray who regulated the furnace draught inversely to steam pressure [6]. His device used the force of steam pressure acting against a weighted piston to drive a damper in the flue gas duct. In 1803 Boulton & Watt used steam pressure to alter the height of water in a column, which, in turn, changed the position of a flue gas damper via a float and chain system [6].
From that time in the early 1800's, while there were some improvements in the hardware used, the application concepts in boiler control did not advance much until the early 20th century [8].
During the early part of this century power stations used only a few absolutely necessary instruments for measuring pressure, vacuum, speed, voltage and current. As additional types of instrumentation became commercially available, more equipment was used to provide data for control and operation of power plant which was consequently growing in complexity [8].
From the 1930's onward, considerable thought was given to automatic control equipment and to the development of automatic controllers for boiler plant operation [9]. Progress was slow at first, because there was much debate about the real need for such equipment, but improvements in instrumentation since the Second World War gave an impetus to the acceptance of automatic control systems. By approximately 1950, boiler control developed into integrated systems for feed water control, combustion control, and steam temperature control [9].
On the plant side, economic considerations have demanded larger and more complex generating units. Correspondingly, the instrumentation requirements have had to keep in step with this development by the provision of more sophisticated automatic control. In the period 1950 to
1970 the development of boiler control was primarily hardware-oriented where many improvements to pneumatic and electronic controllers were made. This further development of controllers, mechanisms, electronics, and relays led to the design of equipment for complete automatic boiler control, and subsequently to schemes for automatic start-up, loading, running 4
and shutting-down of large complicated boiler-turbine units [9].
Historically, meters, gauges, and lights displayed equipment status to the operator, while recorders made a permanent record of plant performance. Remotely operated air cylinders and electric motors served as actuators and gave plant operators the capability of responding quickly and efficiently to changing plant requirements. From 1970 onwards, the development of microprocessors has sparked a beneficial transition to the greater precision of digital control. computer monitors have replaced the panel-board instrumentation, to provide the operator with past and present process information through sophisticated microprocessor-based distributed control hardware [10].
As power plant control became increasingly more complex, the number of measurement signals from the plant, and control signals to the plant has increased too. Currently around 2 000 analog signals and 6 000 binary signals are being installed on a new boiler-turbine unit. There is a gradual movement towards the use of microprocessor-based "intelligent" instrumentation, where, in addition to measuring one or more process variables, self-diagnostics, time stamping, some administrative functions, linearization and even control are also performed by the measuring devices [11]. These instruments are linked to the control system via a two-wire digital bus which conforms to one of a few industrial field bus standards [12].
Today, virtually all control functions are performed digitally by microprocessor-based; programmable controllers. Traditionally, binary control would be done via a programmable logic controller (PLC) while analog control would be done via a distributed control syslem (DCS), but nowadays this distinction is not as clear, and most PLCs and DCSs can do both binary and analog processing [13]. Control algorithms with increased flexibility are becoming available to provide on-line gain scheduling, nonlinear control, instrumentation and actuator linearization, automatic tuning, and many other features [14].
Progress is also being made on advanced control philosophies in many directions. A good example of this is steam temperature control which is one of the most difficult processes to control in steam generating plant. Many different control strategies have been proposed for, and 5
were tested on the steam temperature control loop. This thesis will discuss the various areas of
progress on advanced steam temperature control at a later stage. It will also introduce a new
control philosophy, discuss its advantages and disadvantages and document results obtained on a live 686 MW power plant boiler.
The modelling, practical work and experimentation discussed here was done on Unit 3 at Kendal Power Station, located near Witbank in South Africa. The station comprises six identical boiler-
turbo-generator units, each rated for 686 MW continuous operation. The peak generating
capacity of the station is 4320 MW (6 * 720 MW peak), which rates it as one of the largest coal
fired power stations in the world. •
1.3 The need for steam temperature regulation
In any modem thermal power station, it is of the greatest importance to keep very close control over the steam temperature and temperature gradients, for the following reasons:
Since the expansion of turbine components is directly related to the temperature of steam,
strict requirements on the regulation of steam temperature are imposed by the small
clearances between stationary and moving turbine parts [15].
To maximize the time-to-rupture of boiler components by limiting excessive creep due to
high temperatures [16]. Creep is the time dependent deformation of a material subjected
to stress lower than its yield stress. The creep rate of steel increases with temperature
[17].
To maintain safety margins. A drastic reduction in the yield strength and tensile strength
of steels occurs at temperatures above 540-560°C, depending on the composition of the
steel [17] & [18].
Close matching of steam temperatures to metal temperatures are necessary, especially during start-up and shut-down to prevent distortion on turbine casings [19].
Steam temperature gradients must be kept within tolerances to prevent excessive stress
in the thick-walled components [20]. Repeated temperature transients of an excessive
nature cause thermal fatigue of boiler components.
0 Because the efficiency of the steam cycle is dependent (amongst others) on steam temperature [21], it is beneficial to operate with temperatures as close to the upper limits 6
as possible.
The list above is probably not exhaustive, but it does point out the importance of good steam temperature control on power plant.
1.4 Research hypothesis The following hypotheses underline the work undertaken in this thesis. The heat transfer rate from the firing system to the evaporator, superheater and reheater in a power plant boiler can be modelled by using a neural network trained on real plant test data. Such a neural network model can be used to predict the effect that firing system disturbances will have on the heat transfer rates before the steam temperature is affected significantly by these disturbances. Adjustments to the firing system for minimizing the errors between actual and design heat rates can be obtained by iteratively backpropagating the errors through the neural network. In this way, the effect of firing system disturbances on steam temperature can be largely neutralized.
1.5 Overview of thesis Chapter 2 describes the power plant thermodynamic cycle and defines the various mechanisms of heat transfer between fuel and boiler tubes. It also describes how heat transfer changes with varying boiler load and boiler conditions. The placement and surface area of boiler components and the sensitivity of heated elements to changes in heat distribution patterns are discussed.
Chapter 3 deals with various methods of, and control elements for, steam temperature control. Three main classes of steam temperature control elements are discussed. The effect on steam temperature regulation of long process time lags, variations in process parameters, and process disturbances are presented. The results of a study into the origin . of temperature excursions at Kendal power station are documented. The instrumentation 7 and control configurations applied in practice are discussed and an overview of documented developments in advanced steam temperature control on power plant boilers are made.
Chapter 4 discusses the suitability of applying neural networks to process modelling and control. The artificial neural network, and aspects related to the topology and training of networks, are discussed. Arguments are presented for applying neural networks to the modelling of existing processes. Various neural network controller designs are described, and the error backpropagation technique is shown to be well suited to the steam temperature control problem.
Chapter 5 focusses on the creation and testing of a boiler heat distribution model. The desired characteristics of a heat distribution model for a power plant boiler are listed. The design and execution of a series of live plant tests for acquisition of modelling data are described. Processing the data and calculating the heat transfer rates to the boiler components are described, assumptions are motivated, and the calculation of any unmeasured variables are explained. Specific attention is given to discrepancies in the results. The task of selecting the ideal network topology is described and comparative results are given. Different model output schemes are introduced.
Chapter 6 deals with the design of a neural network based heat distribution controller.
The requirements for improving steam temperature control are listed and it is shown that neural networks lend themselves very well to these requirements. The philosophy of optimal heat distribution (OHD) control is introduced. It is shown how the error backpropagation technique can be applied to calculate optimal control actions.
Chapter 7 describes the implementation and testing of the OHD controller. The development of the software programme and hardware interface is described and intricacies are pointed out. Problems with mill production rates and process noise are addressed. Transient tests are described, and problems experienced with process gain changes, oscillations, and erroneous fuel flow measurements are explained. Final results 8 with OF-ID control are compared to normal P1D control and improvements, and drawbacks, are discussed. 9
2. The power plant boiler
2.1 Cycle description
2.1.1 Carnot cycle In 1824, Sadi Carnot, a French engineer, published a small, moderately technical book, Reflections on the Motive Power of Fire' [22]. With this, Camot made three important contributions: the concept of reversibility, the concept of a cycle, and the specification of a heat engine producing maximum work when operating cyclically between two heat reservoirs each at a fixed temperature. The importance of the Carnot Cycle here is that it forms the basis of the water-steam cycle in power generation.
Figure 2.1 Carnot cycle.
Camot cycles consist of two reversible isothermal and two reversible iserifropic processes (Figure 2.1). A high temperature heat source and low temperature heat sink are placed in contact with the Carnot device to accomplish the required isothermal heat addition
.Q, (a- b) and rejection Q2 (c-d) respectively. The reversible adiabatic process involves expansion that produces work output Wr (b-c) and compression that requires work input We (d-a). The state changes experienced by the working fluid are shown in the temperature-entropy diagram of Figure 2.2.
Translated to English from: Reflexions sur In puissance motrice du feu.
10
T
S Figure 2.2 Carnot cycle T-S diagram.
The classic Camot cycle is such, that no other can have a better efficiency than the Camot value between the specified temperature limits [21]. Other cycles may equal it, but none can exceed it. Practical attempts to attain the Carnot cycle encounter irreversibilities in the form of finite temperature differences during the heat transfer processes and fluid friction during work transfer processes. Moreover, as all of the process fluid has not yet condensed at state d, the compression process (d-a), is difficult to perform on this two- phase mixture. Compressing the gaseous state also consumes large quantities of energy. Consequently, other cycles appear more attractive as practical models.
2.1.2 Rankine Cycle The cornerstone of the modem steam power plant is a modification of the Camot cycle proposed by W.J.M. Rankine [23], a Scottish engineering professor of thermodynamics and applied mechanics. The elements comprising the Rankine cycle are the same as those appearing in Figure 2.1 with the following exceptions: the condensation process accompanying the heat rejection process continues until the saturated liquid state is reached and a simple liquid pump replaces the two-phase compressor. 11
Figure 2.3 Rankine cycle.
Figure 2.3 shows the component layout of the Rankine cycle with a boiler as high temperature heat source, a condenser as low temperature heat sink and a liquid pump replacing the two-phase compressor. The temperature-entropy diagram of the Rankine cycle (Figure 2.4) illustrates the state changes for the Rankine cycle. With the exception that compression terminates at boiling pressure (state a), rather than the boiling temperature (state a), the cycle resembles a Carnot cycle. The lower pressure at state a, compared to a', greatly reduces the work of compression between d-a.
Figure 2.4 Rankine cycle T-S diagram. 12
This Rankine cycle eliminates the two-phase vapour compression process, reduces compression work to a negligible amount, and makes the Rankine cycle less sensitive than the Carnot cycle to the irreversibilities bound to occur in an actual plant. As a result, when compared with a Carnot cycle operating between the same temperature limits and with realistic component efficiencies, the Rankine cycle has a larger net work output per unit mass of fluid circulated, smaller size and lower cost of equipment.
2.1.3 Superheat cycle The turbine in an unmodified Rankine cycle receives dry, saturated vapour from the boiler. Therefore, part of the vapour condenses as it expands and cools through the turbine. In superheat cycles, the vapour is heated above the dry-saturation point, before being fed to the turbine. The use of superheat offers a simple way to improve the thermal efficiency of the basic Rankine cycle and reduce vapour moisture content to acceptable levels in the low-pressure stages of the turbine [21].
Figure 2.5 Superheat cycle T-S diagram.
2.1.4 Reheat cycle Even with the continued increase of steam temperatures and pressures to achieve better cycle efficiency, in some situations attainable superheat temperatures are insufficient to prevent excessive moisture from forming in the low-pressure turbine stages. The solution to this problem is to interrupt the expansion process, remove the vapour for reheating at constant pressure, and return it to the turbine for continued expansion to condenser 13 pressure (Figure 2.6). The thermodynamic cycle using this modification of the Rankine cycle is called the reheat cycle. Reheating may be carried out in a section of the same boiler supplying primary steam, in a separately fired heat exchanger, or in a steam-to-steam heat exchanger. Most present-day utility units combine superheater and reheater in the same boiler [4].
Figure 2.6 Reheat cycle with economizer.
For large installations, reheat makes possible an improvement of approximately 5 percent in thermal efficiency and substantially reduces the heat rejected to the condenser cooling water [24]. The operating characteristics and economics of modern plants justify the installation of only one stage of reheat except for units operating at supercritical pressure. One further addition to the Rankine cycle for increasing efficiency Was that of the economizer. This element raises the temperature of feed water by utilizing the low temperature heat after the flue gas had been cooled by evaporator, superheater and reheater (Figure 2.7).
14
T Superheater Evaporator 12\ eheater
Economizer Allik-4 1
Feed pump Turbines
Condenser
Figure 2.7 Reheat cycle with economizer T-S diagram.
2.1.5 Regenerative Rankine cycle Refinements in component design soon brought power plants based on the Rankine cycle to their peak thermal efficiencies, with further increases realized by superheating and reheating the steam as described above. Efficiencies were further boosted by increasing the temperature of the steam supplied to the turbine and by reducing the sink (condenser) temperature. Currently, all of these are employed with still another modification, being regeneration.
The regenerative cycle reduces irreversibility by bleeding hot, partially expanded steam from the turbine(s) and using it to heat the compressed water fed to the boiler. In this way it increases the overall cycle efficiency. Apart from increasing cycle efficiency, regeneration impacts the process in two ways: it changes the temperature of the boiler feed water and it reduces the steam flow through the reheater. These two issues will be discussed in more detail later in Chapter 5.
2.2 Heat transfer theory During the combustion process inside a furnace, enormous quantities of chemical energy is converted to heat and discharged into the furnace space. Most of this heat is transferred to the boiler tubes and working fluid while a small percentage is lost to atmosphere through the hot flue gas. Heat transfer takes place through three individual mechanisms: conduction, convection and radiation. In a power plant boiler, heat is transferred simultaneously by all three mechanisms. 15
The mechanisms of heat transfer will be discussed here to point out the factors influencing heat transfer between the burning fuel and the working fluid. For the purpose of this thesis it is not necessary to do an in-depth analysis of heat transfer. However, it is important to emphasize the differences in the physical mechanisms of heat transfer and to discuss the main factors influencing it
2.2.1 Conductive heat transfer
Conduction takes place by elastic molecular impact, molecular vibration and in metals by
electronic movement. In comparison to heat transfer through convection and radiation,
heat transfer by means of conduction through the flue gas to the boiler surfaces is
negligibly small [25]. However, heat conduction theory does play a role at the boiler tube
surface where the heat has to pass through the metal tube wall or through a covering layer
of ash or slag.
The equation for heat conduction through multi-layer cylindrical walls [26] can be written
to apply to heat conduction through a boiler tube covered with ash:
2n-L(Tg - Tf) qcond In(r olr ,) In(r jr 0) (2.1) Ict Ica
where:
q gond heat transfer rate through boiler tube and ash [W] thermal conductivity of the boiler tube metal [W/cm°C]
ka = thermal conductivity of ash [W/ m°C]
Tg = temperature of combustion gas [°C] = 7} temperature of fluid inside boiler tube [°C] L = length of the boiler tube [m] r, = inner radius of boiler tube [m] ro = outer radius of boiler tube [m]
ro = outer radius of ash layer [m]
The thermal conductivity lc, of steel ranges between 20 and 50 W/ m°C depending on its 16
temperature and composition [26]. Much lower is the thermal conductivity of ash and
slag, both being below 1.0 W/ m°C [26]. Therefore, if an ash layer forms on a boiler tube, it significantly reduces, and quickly dominates, the heat transfer rate into the tube.
Due to this reduction in heat transfer, modern furnaces have high pressure sootblowers installed to periodically blow the contaminants from the heat transfer surfaces.
2.2.2 Convective heat transfer
Convection in a power plant boiler involves transportation and exchange of heat due to
-flue gas motion and is governed by the laws of aerodynamics and fluid dynamics.
Convective heat transfer is described by Newton's law of convection [26]:
(2.2) q cony = kconv A (T. -
where:
qcond = convective heat transfer rate to boiler tube [W] A = surface area of the boiler tube [m 2] k = convection heat transfer coefficient [win12..c] To, = temperature of the free gas stream [°C] Tw = surface temperature of convector [°C]
The convection heat transfer coefficient is sometimes called the film conductance because of its relation to the conduction process in the stationary layer of fluid at the wall surface.
The convective heat transfer coefficient is dependent on numerous gas .property and dimension related variables. Singer [4] states the following expression for film conductance: Rc = f (D, V, p, ,u, ci,, k, a) (2.3) where: R, = film conductance [W/m2 °C]
dimension of boiler tube surface parallel to gas flow [m]
Vg = linear velocity of gas stream [m/s]
Pg density of flue gas [kg/m3]
/-18 = viscosity of flue gas [kg/m.s] specific heat of flue gas at constant pressure [J/kg°C] Pt 17
kg = thermal conductivity of flue gas [W/m°C] a = geometric relation ratio to cover the effect of tube spacing, width, depth and length [dimensionless]
Of the seven parameters affecting film conductance, D and a remains constant for a given boiler, while pg, pg, cgg and kg change only a few percent with flue gas temperature and composition (see Table 2.1). On the other hand, the flue gas velocity V, may change through an order of magnitude from minimum to maximum boiler load, since furnace air flow varies proportionally to furnace fuel flow. The value of Rc changes from 6.5 to 180 W/m2 °C as air flow around a 50 mm diameter horizontal tube increases from natural convection to 50 m/s forced convection [26].
Temperature [°C] pg [kg/m1 pg [kg/m.s] cgg [kJ/kg°C] kg [NV lm°C] 1273 0.3524 4.152 1.1417 0.06752 1773 0.2355 5.400 1.230 0.0946 Table 2.1 Properties of air at atmospheric pressure. [26]
Flue gas velocity is also important from a control perspective: of the seven variables influencing the convective heat transfer, it is the only controllable variable, although within certain limits. This concept will be utilized for control purposes later.
2.2.3 Radiant heat transfer In contrast to the mechanisms of conduction and convection, where heat is transferred through matter, heat may also be transferred through regions where a perfect vacuum exists. Thermodynamic theory shows that an ideal thermal radiator, or blackbody, will emit energy at a rate proportional to the fourth power of the absolute temperature of the body and directly proportional to its surface area [27]. Thus a A T 4 (2.4) g rad = where a is the Stefan-Boltzmann constant and has the value of 5.669 x 10.8 W/m2K4 and T is measured in kelvin. Equation (2.4) is called the Stefan-Boltzmann law of thermal 18
radiation, and it applies only to black bodies. The net radiant exchange between two
surfaces will be proportional to the difference in absolute temperatures to the fourth power, i.e., (2.5) grad cc A ° ( 7.14 T24)
Boiler heat transfer surfaces are generally not black, but are covered with a layer of dark gray iron oxide or gray ash. To take account of the gray nature of boiler surfaces, another factor is introduced, called the emissivity e. This factor relates the radiation of a gray
surface to that of an ideal black surface.
(2.6) g rad = E A a (7:14 -
The emissivity of boiler surfaces depends on the cleanliness thereof and the colour and composition of the iron oxides and ash, but generally c = 0.762 [25]. The interpretation of Equation (2.6) is that radiant heat transfer will vary proportional to the fourth power of flame temperature and air flow rate has no direct effect on it.
2.2.4 The effect of nonluminous radiation
Carbon dioxide and water vapour are the principal radiating components of boiler flue gas
[4]. Their combined radiating effect has historically been referred to as nonluminous radiation. In all cases where the flue gas temperature is high and the tube spacing relatively large, the nonluminous radiation will be of considerable magnitude.
Nonluminous radiation is proportional to the difference in temperature between the flue gas and the boiler tubes and, therefore, its effect can be added to the convection rate [4].
2.2.5 Total heat transfer
The total rate of heat transfer between the furnace flame and boiler components is a complex combination of the basic equations given above. Deriving the total heat transfer from first principles lies beyond the scope of this thesis and the reader is referred to [28]
& [29] for a complete discussion of the subject. 19
2.3 Steam generator design A steam generating unit may be considered to have two sections: one is responsible for generating heat (the furnace, or fire side) and the other absorbs the heat (the boiler, or water side). The boiler consists mainly of tubes and it encloses the furnace. The furnace consists mainly of empty space for combustion, but the burners are also considered to be part of the furnace. Sometimes, the term boiler is used when referring to the entire steam generating unit, including the furnace.
2.3.1 Fireside (furnace) In the process of steam generation, fuel burning systems provide controlled, efficient conversion of chemical energy of fuel into heat energy which, in turn, is transferred to the heat absorbing surfaces of the steam generator. To do this, the fuel burning system introduces fuel and air for combustion into a furnace, mix and ignite these reactants, and distribute the flame envelope and products of combustion.
Figure 2.8 Fire-side components of a steam generator.
The basic power plant furnace is a hollow chamber into which fuel and air is introduced for combustion (Figure 2.8). In the case of coal fired furnaces, technology has progressed from moving bed furnaces burning crushed coal to pulverised fuel systems burning fine coal powder [4]. In these systems coal is pulverized in mills (also called pulverizers) and transported to the furnace by blowing it from the mills along fuel pipes by means of an air 20 supply called primary air. The primary air needed for transportation is only about 15-20% of the total air required for combustion, hence the addition of secondary air at the burner nozzle [29].
Power boilers are designed with 4 to 6 mills, each mill feeding 4 to 8 burner nozzles. Firing systems are mainly classified as horizontally wall-fired systems (characterized by individual flames), tangentially fired systems (which have a single flame envelope) and vertically fired systems (which have individual flames merging into one flame envelope) [4 . The different firing systems are shown in Figure 2.9.
-4t 4-
a. n. c.
Figure 2.9 Different firing systems indicating fuel injection angle: a) Horizontally fired, b) Tangentially fired - top view, c) Vertically fired.
Horizontally Fired Systems In this design, the coal and primary air are introduced tangentially to the burner nozzle, thus imparting strong rotation within the nozzle. Adjustable inlet vanes impart a rotation
to the preheated secondary air from the windbox. The degree of air swirl , coupled with the flow-shaping contour of the burner throat, establishes a recirculation pattern extending several throat diameters into the furnace. Once the coal is ignited, the hot products of combustion propagate back toward the nozzle to provide the ignition energy necessary for stable combustion. The burners are located in rows, either on the front wall only or on both front and rear walls. The latter is called "opposed firing." In general, each row of burners will be served by a different mill [4].
Tangentially Fired Systems The tangentially fired system is based on the concept of a single flame envelope. Fuel and 21 secondary air are projected from the corners of the furnace along a line tangent to a small circle, lying in a horizontal plane, at the centre of the furnace. Intensive mixing occurs where these streams meet [4]. A rotating motion, similar to that of a cyclone, is imparted to the flame body, which spreads out and fills the furnace area. As with horizontally fired systems, the burners are located in rows, with each row being served by a different mill.
When a tangentially fired system projects a stream of pulverized coal and air into a furnace, the turbulence and mixing that take place along its path are low compared to horizontally fired systems. This -occurs because the turbulent zone does not continue for any great distance, since the expanding gas soon forces a streamline flow. However, as one stream impinges on another in the centre of the fintace, during the intermediate stages of combustion, it creates a high degree of turbulence for effective mixing. This creates a "fireball" effect where fuel from individual mills is discharged into a high intensity heat envelope [4].
Vertically Fired Systems The first pulverized coal systems had a configuration called vertical, down-fired or arch firing. Pulverized coal is discharged vertically downward through burner nozzles located on extension surfaces on two sides of the furnace. The firing system produces a long, looping flame in the lower furnace, with the hot gases discharging up the centre. A portion of the total combustion air is withheld from the fuel stream until it projects well down into the furnace [4]. This arrangement is less common in large power boilers and will not be treated further here.
2.3.2 Water-side (boiler) The water-side of a steam generator comprises the economizer, evaporator, superheater and reheater. Water is admitted to the boiler and passes through the economizer where it is heated close to, but below boiling point. From the economizer the water is passed to the evaporator where it is boiled to steam. The steam is separated from the water and passed through the superheater where its temperature is increased to the nominal turbine inlet design temperature (actually, the temperature of the steam leaving the superheater 22
is slightly above turbine inlet design conditions to offset the temperature decrease through
the main steam pipes). Once the steam has passed through the high pressure turbine it is
readmitted to the boiler where its temperature is again raised in the reheater. The reheated
steam is passed through the intermediate and low pressure turbines after which it is condensed back to water in the condenser.
High Pressure Turbine Low Pressure Turbine Saturated Water —111=C Wate Saturated Steam 1—/IIC, —0.- SHS SHS To Condenser a a
SHS = Superheated Steam
Economizer Evaporator Superheater Reheater
Figure 2.10 Diagrammatic view of the water & steam path through power plant components.
2.3.3 Boiler heat transfer surface design
The calculation of boiler heat transfer area presents a great challenge to boiler design engineers. Not only does 'the design have to absorb the maximum possible quantity of available heat, but it has to do this at the lowest possible cost. The boiler has to maintain a maximum efficiency throughout its design range. This calls for a carefully calculated balance between the radiant and convective heat transfer surface. Although much theory has been developed around the mechanics of heat transfer (for exampl6 [30] & [31]), boiler manufacturers rely largely on operational experience backed up by scientific data
[29], and computer simulations [32] when designing heat transfer surfaces in boilers.
One of the most pronounced phenomena influencing the balance between convective and radiant boiler surface, is that radiant heat transfer does not increase as rapidly as convective heat transfer with increasing boiler load [33]. The increase in furnace draught in a sense cools down the combustion process while it increases gas velocities. Therefore, the flame temperature does not increase much with load [34]. Consequently, a larger increase in convective heat transfer occurs through loading than the increase in radiant 23 heat transfer (Figure 2.11).
ture era temp Steam
40 60 80 100 % Steam flow
Radiant superheater Convective superheater Superheaters in series
Figure 2.11 Typical steam temperature characteristics. [28]
Boiler surface design needs to take this into account by finding the best balance between convective and radiant surface throughout the boiler load range. The balance must be maintained when firing any fuel that has been specified for the boiler, and under varying load conditions. It may also be noted that the proportioning of heat distribution varies with the cycle pressure. This is illustrated in Figure 2.12.
At first sight of a sectional side elevation of a modem power boiler it may seem that although the gas flow is quite simple, the water and steam flow path is unduly complicated or even random. But in fact, the disposition of the various parts of the cooling surface is carefully considered to make the most economic use of natural, physical heat transfer phenomena. It is possible to classify the heat transfer space into three main zones: radiation zone, convection zone and heat recovery zone [29]. The approximate borders of these zones are shown in Figure 2.13. 24
3500 r ASuperheater heat rise _3000
2500 Evaporator heat rise a 2000 .c
1500
1000 I 10 12 14 16
Economizer discharge Evaporator discharge Superheater discharge
Figure 2.12 Heat rise in boiler elements vs. steam pressure.
The radiation zone This is the furnace combustion zone of the steam generator. Here radiation and the high temperature gas of combustion is be used for heating water and steam with a low to medium degree of superheat [29]. The temperature of the gas where it leaves the 'radiation zone is referred to as the furnace exit temperature. The convection zone Here medium temperature gas can be used for heating steam with a medium to high degree of superheat [29]. The final stages of the superheater and reheater are normally positioned at the start of the convective zone. The heat recovery zone This zone is situated in the boiler backpass. With cooler flue gas, heat can only be absorbed effectively by cool fluids, such as feed water and steam with a low degree of superheat [29]. It is therefore a favourable location for the initial stages of the superheater and reheater. Also, towards the boiler exit, where the gas has cooled down significantly, one finds the economizer. 25
Figure 2.13 Different heat zones in a steam generator. [29]
Within these zones there is scope for placement of superheater and reheater surfaces allowing the designer to provide for absorption of the correct proportion of heat in all the boiler stages as well as to provide for the correct total heat absorption.
2.3.4 Heat transfer requirements of boiler elements
Evaporator Heat generated in the combustion process appears as furnace radiation and sensible heat in the products of combustion. Most modern boiler have integral furnaces enclosed by water filled wall tubes that serve as the evaporator [28]. By enclosing the furnace, the evaporator receives most of the available radiant heat. Water circulating through the wall tubes absorbs around 50 percent (this will be shown later) of the total heat discharged, and generates steam through the evaporation of part of the circulated water. The absorption of such a large portion of the heat of combustion serves to reduce the temperature of the . gas entering the convective zone to the point where slag deposit can be controlled by soot blowers [29]. Utilizing radiant heat discharge for evaporation is convenient from a thermodynamic point-of-view, because as the ratio of radiant heat transfer to steam flow 26
decreases with boiler load (Figure 2.11), so does the heat needed for evaporation (Figure 2.12).
Superheaters And Reheaters As discussed earlier, the function of a superheater is to raise the boiler steam temperature above the saturated temperature level. As steam enters the superheater in an essentially dry condition, further absorption of heat sensibly increases the steam temperature. The reheater receives superheated steam which has partly expanded through the turbine and re-superheats (reheats) this steam to a desired temperature.
Superheater and reheater design depends on the specific duty to be performed. For relatively low final outlet temperatures, superheaters solely of the convection type are generally used [4]. Towards the end of the convective zone, horizontal tube banks are installed as low temperature superheater or reheater sections. The boiler roof and backpass walls are covered with low temperature superheater panels, also for convective heat transfer.
For higher final temperatures, surface requirements are larger and, of necessity, superheater elements are located in radiation and'very high temperature convective zones. Radiant wall type superheaters and reheaters and widely spaced tube panels (located on horizontal centres of 1.5 m to 2.5 m) allow substantial radiant heat absorption [4]. Platen sections (tubes separated with steel plate strips to form a solid plate-like bank, on 0.35 m to 0.7 m centres) are placed downstream of the panel sections to provide high heat absorption by both radiation and convection [4].
Economizers Economizers help to improve boiler efficiency by extracting heat from low temperature flue gas after the convective zone. The economizer heats feed water, which enters at a temperature appreciably lower than that of saturated steam. Due to its low inlet and discharge temperatures, economizers are suitably located in the cooler heat recovery zones [4]. 27
Air heaters Air heaters do not form part of the water-side of a steam generator, but because it forms part of the heat recovery equipment, it is mentioned here for the sake of completeness.
Steam generator air heaters cool the flue Eps before it passes to the atmosphere while they raise the temperature of the incoming air of combustion, thereby increasing fuel firing efficiency. In theory, only the primary air (used to dry the coal in the mills) must be heated. Ignited fuel can burn without preheating the secondary air [4], but there is considerable advantage to the furnace heat transfer process in heating all the combustion air: it increases the rate of burning, helps raise the flame temperature and increases boiler efficiency. Air heaters are located below the backpass, the furthest away from the furnace, ending off the heat recovery zone.
/ Steam cooled roof \\
Pendant convection superheater or reheater Radiant wall Horizontal reheater convection superheater Zor reheater Panel type superheater Superheater steam —3o- Economizer cooled walls 7 Platen type superheater Or reheater
Air heater ••• •• • •• • Furnace walls • • •
Figure 2.14 Typical location of steam generator elements. [4]
Figure 2.14 shows the typical placement of heat absorbing elements within a modem power boiler. 28
2.3.5 The Kendal Boiler
The boilers at Kendal Power Station were designed by Combustion Engineering (now incorporated into ABB). All the boilers are rated for a maximum main steam flow of
577 kg/s at 540 °C and 16.5 MPa. The final reheat steam temperature is also 540 °C.
The furnaces are of the tangential, corner fired type. Each boiler has five ball mills providing pulverized coal fuel for combustion. Every mill serves a different elevation of eight burner nozzles, two per boiler corner.
These boilers deviate from the standard Combustion Engineering design in two areas: vertical burner spacing and a reheater with mainly convective heat transfer surface [35].
Vertical Burner Spacing Based on experience with slagging on units which had a firing zone heat release rate which was too high, Eskom specified a maximum furnace heat release rate of 1 MW / nf. The final boiler design involved a conservatively sized furnace and a firing system with increased vertical spacing between burner levels [35]. A typical 550 MW boiler of similar design (Arnot Power Station) has a distance of 8.2 m between its lowest and highest burner nozzles, while the 686 MW Kendal boilers have a distance of 23.6 m here [36].
The large distance between burner elevations at Kendal results in a noticeable difference in heat transfer pattern depending on which mills are in service at any time (this will be shown later).
Convective Reheater On units without an Hp turbine bypass system, furnace temperatures must be carefully controlled prior to admission of steam to the turbine because there is no reheat steam flow to cool the radiant reheater tubes. This is especially critical for a radiant reheater.
Although the Kendal units were specified to have HP bypass systems, Eskom specified that the boiler not have a reheat radiant wall. Eskom did not want the operators to deal with the consideration of furnace temperatures during the unusual startups when the bypass would not be available for some reason [35].
▪ ▪
29
These wishes were accommodated by designing a virtually 100% convective reheater and balancing the surface by using a radiant wall superheater in addition to the predominately radiant superheater division panels [35]. Due to its mainly convective nature, the Kendal reheaters are very sensitive to the furnace air flow rate. Additionally, due to the lack of radiant surface, the design reheat steam temperatures cannot be maintained under low load conditions.
The placement of heat transfer surface area in the Kendal boilers is shown in Figure 2.15.
In comparison to a standard Combustion Engineering boiler, Figure 2.14, the Kendal boilers have more radiant superheater surface while having virtually no radiant reheater surface.
Boler mot perimeter \_A
Pendant ccovection reheater Horizontal Rivfont convection superheater reheater
Divisional panel superheater
Platen type Back pass walls Economizer hiah temperature superheater superheater Pendant type low ternperanere superheater
Burner nozzle eleventh's
Furnace vigilsvigils \eapciatcr/
Figure 2.15 Layout of the Kendal boiler heat transfer elements. 30
3. Steam temperature control
3.1 Control elements for steam temperature regulation As described in the previous chapter, heat transfer to the superheater and reheater is a function of many variable process parameters. The necessity of keeping the steam temperatures as close fo design as possible was also stated earlier. Consequently, the boiler designer has to allow for some means of influencing the steam temperature in order to compensate for any process fluctuations that can change the steam temperature.
The options available to the designer are: changing the combustion gas temperature, or its mass flow rate, or changing the steam mass flow rate or reduce its enthalpy. Steam temperature control devices are incorporated in the boiler firing system, in the superheater or reheater circuitry, or in arrangements of dampers for gas bypass. The following means of steam temperature control are applied, [4], [8], [28], [29], [37], [38]:
• Desuperheating by water sprayed into piping ahead of, in between, or following superheater or reheater sections. • Firing system manipulation in which the effective release of heat from the fuel burning process is made to occur at a higher or lower portion of the furnace. This affects the heat absorption pattern in the furnace and, consequently, the radiation zone exit gas temperature. • Recirculation of gas, in which a portion of the combustion gases are brought back to the furnace and are added to the normal once-through flow of gas passing dyer superheater and reheater. • Gas bypass around some of the installed heating surface that provides excessive heat in certain parts of the load range. The purpose is to preVent such surfaces from absorbing heat from the bypassed gas so that the desired steam temperature is achieved without using any other means. Excess air concentration influences the balance in heat transfer between radiant and convective surfaces. • Selective soot blowing reduces heat transfer to elements by letting them foul up with ash and slag. 31
• Utilizing a separately fired superheater allows independent temperature control by means of firing rate manipulation.
The following few subsections describe in more detail these different methods of control, used in one form or another by all manufacturers.
3.1.1 Desuperheating
Desuperheating is the reduction of temperature of superheated steam accomplished by spraying water into the piping or by diverting steam flow through a heat exchanger for
cooling. The desuperheating Water must be of very high purity and may be supplied from
the feed water line [28]. The heat exchanger-type desuperheater uses boiler water as the
cooling medium, either by diverting it through an external heat exchanger [29] or by
diverting superheated steam through heat exchanger tubes integral to the boiler drum [28].
Many large boiler installations use desuperheating in combination with one or more of the
other temperature control methods [4]. If desuperheating is to be the only method of
steam temperature control on a specific boiler, the heated elements must be designed with
excessive heat transfer surface. Consequently, the steam temperature will be excessively
high and a desuperheater can be used to: remove this excess temperature [4].
Desuperheating of reheat steam is generally not desirable because of its adverse effect on
plant efficiency: the water used for desuperheating has bypassed the entire high pressure
cycle. Consequently, reheat outlet temperature is best controlled using some means other
than water spray, unless it is unavoidable [28].
If located beyond the outlet of the superheater, a desuperheater will condition the steam
before it is passed along to the turbine. Although this arrangement may be practical for
low temperature superheaters, the preferred location of the desuperheater is between
sections of the superheater [4]. In such interstage installations, the steam is first passed
through one or more primary superheating sections, where it is raised to some
intermediate temperature. It is then passed through the desuperheater and its temperature
controlled so that, after continuing through the secondary or final stage of superheating, 32
the required constant outlet temperature is maintained.
The heat given up by the steam during a temperature reduction is picked up by the cooling water in three steps. First, its temperature is raised to that of saturated water, then the water is evaporated, and finally, the temperature of the steam so generated is raised to the final condition of temperature at the desuperheater outlet. By setting up a simple heat balance equation, it is possible to determine exactly the quantity of water required to desuperheat for any given set of conditions. It will be shown later how the method of heat balance across a desuperheater was applied in practice.
Desuperheating can only lower the temperature of steam. If it is necessary to also raise the steam temperature, other methods, such as those discussed below, must be incorporated into the boiler design.
3.1.2 Firing system manipulation There are two common ways to vertically displace the zone of highest heat release in a furnace to achieve a change in the outlet gas temperature [33]. The first, often used with wall fired, fixed burners, is to insert or withdraw levels of burners as a function of load [28]. Removing lower levels and firing through the remaining upper levels effectively moves the heat release zone higher in the furnace. Because continuous (analog) control
is not possible in this way, it necessitates backup by spray desuperheating for vernier . control.
Tilting fuel and air nozzles, used in corner (tangential) fired systems is a practical method of controlling furnace outlet gas temperature smoothly without cycling equipment in and out of service [4]. Depending on design, superheater or reheater steam temperatures can be regulated by changes in burner nozzle tilt angle. 33
R t \ / E
A
( - - )11. • (- ) I. .4 ' L •••• Burner angle Burner angle - S1 Burner angle • = +30 deg = 0 deg = -30 deg . .
Figure 3.1 The effect of burner tilt angle on fireball elevation. [4]
The adjustment of the burner tilt angle alters the position of the fireball within the furnace (Figure 3.1) and hence alters the furnace heat absorption [37]. The gas temperature leaving the furnace for a given fuel flow rate is directly related to the furnace heat absorption and hence to the burner tilt angle (Figure 3.2).
1300
'&1200 E
g 1 150 C
LL
1100 I i I I 1 I -30 -20 -10 0 10 20 30, Burner tilt angle [deg]
Figure 3.2 Effect of burner tilt angle on furnace exit temperatures. [33] 34
The main effect of the variation of the tilt angle is to alter the rate of heat absorbed by the high temperature surfaces situated immediately beyond the furnace [4]. Directing the flame toward the upper part of the furnace maintains a higher gas outlet temperature than is the case if the flame were directed horizontally into the furnace. Burners may be tilted upward during low load conditions or when the furnace walls are clean. At higher loads, or when the walls are coated with ash or slag, burner nozzles can be positioned horizontally or angled downward to decrease the furnace exit temperature [4]. A shortfall of tilting burners is that the buoyancy of the hot furnace gas tend to make tilts below -15° less effective. Mother disadvantage is that the burner boxes are prone to seizure and loose their effectiveness in steam temperature control [37].
A third method of manipulating the firing system is to bias the fuel flow rate at different elevations. (This method is believed to be quite uncommon - of nine references discussing steam temperature control methods, only one reference, [38], briefly mentions mill biassing.) The effect of mill biassing is similar to tilting burners or placing burner elevations in and out of service - it positions the heat release area higher or lower in the furnace. This is achieved by firing more fuel through the upper burners than through the lower ones or vice versa.
3.1.3 Flue gas recirculation In this temperature control method, a portion of the combustion gas is diverted from the main stream at a point following the superheater and reheater (usually between the economizer outlet and the air heater inlet [4] or after the economizer [37]) and is recirculated to the furnace where it is introduced in the immediate vicinity of the initial burning zone. The gas passes through a recirculating fan and mixes with the gas in the furnace, lowering its temperature and consequently causing a reduction in radiation heat transfer. As a result, the heat available to the superheater and reheater increases, as does the quantity of gas passing over the surfaces which increases convective heat transfer. Both of these factors increase steam temperature [37]. 35
An alternative to gas recirculation, called gas tempering, also diverts gas from the main stream after the economizer, but introduces it near the furnace outlet, before the convective zone [28]. While gas recirculation decreases the furnace radiant heat transfer rate and increases the rate of heat transfer to all the other boiler elements, gas tempering does not alter the heat absorbed by the furnace. It does,.however, reduce the furnace exit temperature while increasing the gas velocities. This has the effect of reducing the heat transfer rate to the radiant superheater and reheater while increasing the heat transfer rate to the convective elements [28].
In both arrangements, the flue gas should have a low ash content to prevent serious abrasion of the recirculation fan impeller. In coal fired boilers, this problem can be overcome by extracting the recirculating gas from after the induced draught fans [37]. At this point, the flue gas has been cleaned from most of the ash by passing through bag filters or electrostatic precipitators. This system has the added advantage that the induced draught fans can be sized to produce the head necessary to recirculate the gas without the need for additional gas recirculation fans.
Flue gas recirculation may be used to supplement "normal" temperature control [38]. For instance, when used in conjunction with fuel nozzle tilt control, gas recirculation may be applied to maintain the fuel nozzles in their horizontal position.
3.1.4 Flue gas bypass
The boiler convection banks can be arranged in such a manner that a poi -lion of the flue gas can be bypassed around some of the superheater elements [28]. The superheater is oversized in design so that it will produce the required degree of superheat at partial load conditions, say 75 %. As the load increases, some of the flue gas bypasses the respective superheater sections.
Although the gas dampers are made of alloy steel, they cannot be installed in a high temperature zone. Gas bypass control is popular because of its low initial cost, but the regulating dampers are difficult to maintain because of the high temperatures to which 36 they are subjected [4].
3.1.5 Excess air flow rate The steam outlet temperature of a convection superheater may be increased by increasing convective heat transfer by increasing the excess air supply [37]. However, the additional gas mass flow will reduce the gas temperature and decrease the radiant furnace heat absorption for a given firing rate. The increased gas mass flow with its increased total heat content serves to increase the degree of convective superheat. Radiant superheaters receive less heat transfer. Unlike the gas recirculation method, an increase in excess air decreases the boiler efficiency because more heat is lost through the smoke stack in terms of excess heated air. However, the stack losses may be offset by an increase in turbine efficiency as a result of higher final steam temperatures [28].
A variation on adjusting the excess air ratio is called air injection [37]. With this scheme, some additional heated combustion air is diverted from the secondary air ducts into the furnace hopper area (below the combustion zone). Except for the point of injection, air injection has the same properties and effects as excess air control.
3.1.6 Selective soot blowing The control of superheat with soot blowers is accomplished as follows [33]: When superheat is low, the radiant superheater surface is cleaned to increase the total heat absorbed by the superheater. When superheat is high, other furnace surfaces are cleaned to increase the effectiveness of the furnace cooling surface and hence reduce the percentage of heat absorbed by the superheater. Selective soot blowing cannot be used for active steam temperature control as its response time is far too long.
3.1.7 Separately fired superheater A superheater, completely separate from the steam generating unit and independently fired may be utilized as an alternative method of controlling superheater outlet temperatures 37
[33]. The degree of superheat is directly influenced by the firing rate onto this separate
superheater. This arrangement is not generally economical for power generation where
a large quantity of superheated steam is needed, and its use is largely confined to process
industries, such as chemical manufacture and petroleum refining [28].
3.1.8 Steam temperature control at Kendal
The Kendal boilers are provided with three mechanisms for controlling the steam
temperature: tiltable burner nozzles, desuperheating spray stations on the superheater and
the reheater and a variable excess air ratio. These will be discussed individually.
Burner tilts
The injection angle of the fuel burners is continuously adjustable through an angle of -30°
to +30°. Because the superheater surface has predominantly radiant surface, the burner
tilts have a significant effect on the heat transfer to the superheater; much more so than
on the convective reheater.
Mill combination Burner tilt angle
ABCDE (all mills in service) 0°
ABCD (E-Mill out of service) _150
ABCE (D-Mill out of service) -7.5° ABDE (C-Mill out of service) 0°
ACDE (B-Mill out of service) +7.5°.:;
BCDE (A-Mill out of service) +15°
All 3-Mill combinations -15°
Table 3.1 Mill combinations and corresponding tilt angles.
During the commissioning of the Kendal Units various control strategies were tried with the burner tilts as final control element. The final control arrangement compensated for various mill combinations and also provided steam temperature control when certain temperature limits were exceeded. Varying the burner tilt angle attempts to keep the 38
furnace heat discharge as central as possible. For example: if the top mill (A-Mill) is out
of service, the burner tilts should be aimed upwards to compensate for the loss of heat high up in the filmace. The setup in Table 3.1 was heuristically arrived at by Combustion
Engineering and Eskom commissioning staff. Aiming the burner tilts to -15° with three- mill combinations was found to assist combustion stability under low loads.
The preselected burner tilt angles in Table 3.1 are overrided by a final steam temperature
below 530°C, which increases the tilt angle, or by too high interstage steam temperatures,
which decreases the tilt angle.
Desuperheating
The Kendal units were originally designed with a single desuperheating stage located
immediately after the primary superheater. The designers anticipated that the conservative
furnace size, coupled with the unique tilting burner capability of the ABB/CE boiler
design, would keep the superheater heat pick up within the spray capability of the single
desuperheater stage [35]. When the first Kendal unit went into service it became clear
that at 50-60% unit load, the ability to control steam temperature at steady state was very
sensitive to which mills were in service. Steam temperature control was completely
unsatisfactory when making load changes in this mid load range. The quantity of spray
which could be introduced was limited by the requirement to maintain the desuperheater
outlet temperature at least 10°C above saturation temperature for good evaporation [35].
Combustion Engineering proposed, and Eskom accepted, the addition of a second stage desuperheater located at the division panel outlet (Figure 3.3). This second desuperheater
station allowed more spray to be used due to the larger margin above saturation at this location. The added temperature control loop could also be tuned faster for improved control response because there is less surface between this location and the superheater outlet.
Superheater steam temperature control was greatly improved by the second desuperheating station, but executing load ramps in the 50-70% load range with the top 39 mills in service still resulted in steam temperatures leaving the left side division panel outlet header exceeding 550°C.
First stage Second stage desuperheater desuperheater
Boiler roof Low temp Radiant Divisional High temp and pendants walls panels platens backpass
First stage Second stage desuperheater desuperheater
Figure 3.3 Kendal superheater stages and desuperheater locations.
Excess Air Because of the mainly convective reheater, furnace air flow may seem a viable method of temperature control. This option was however not pursued as the primary means of
reheat temperature control , as it was feared that a fuel-rich mixture remaining after reducing the air flow may increase the probability of a furnace explosion [36].
The primary means of reheater temperature control is by means of a desuperheater station at each of the two reheater inlets. These provide short term temperature regulation. In a the long term, the quantity of excess air is adjusted through a ratio controller to keep the total desuperheater spray water flow to the reheater equal to 2.5% of the main steam flow. The value of 2.5% was determined practically as being the minimum average quantity of desuperheater flow required for smoothing out temperature deviations.
Excess air is controlled through the 0 2 controller which measures the percentage of free oxygen in the flue gas, and manipulates the furnace draught to keep this 0 2 measurement to its setpoint. In turn, the furnace draught influences the convective heat transfer to the reheater. 40
3.2 Difficulties associated with steam temperature regulation Steam temperature control has historically been considered the most difficult of all boiler control loops to optimize [39], [40]. This is partly due to the number and extent of unmeasurable disturbances influencing the final steam outlet temperatures, and partly due to dead time, long time lags, nonlinearities and process parameters that change over time. Interaction between the temperature control loop and other loops in the boiler control system adds to the complexity of the problem [40]. This section explains the factors bringing about the challenge associated with good steam temperature control in thermal power plant.
3.2.1 Process disturbahces The previous section showed that heat transfer depends on many factors. Some of these factors are fixed by design (e.g. location of heater elements), but others may change during boiler operation and consequently, it may disturb the heat transfer. Changes in heat transfer will affect the final steam temperatures. Below is a list of process disturbances that can affect steam temperatures; Boiler load. A continued constant load is rarely found except perhaps in high-capacity, high efficiency units that are prime loaded while variable loads are handled by other units [39]. To maintain or change boiler load, the fuel firing rate is manipulated to obtain a specific steam pressure at the superheater outlet. Therefore, the firing rate is dependent on boiler load and is not concerned with the steam temperatures. However, as the heat distribution changes through boiler load, so will the steam temperatures (at least until the closed loop pi:introl returns steam temperatures to setpoint). Fuel type. The steam temperature can be affected by a change in fuel type, depending on the luminosity of the flame and the rate of combustion [25]. Taking samples of the coal being burnt at Kendal Power Station showed variations in calorific value of up to 10% in 24 hours. Burner operation. Most power plants are capable of delivering full load with one or two pulverisers out of service [28]. This is a requirement to ensure that the maintenance of pulverisers does not impose plant load losses. If the upper burners are in service; the furnace exit temperature is higher than with the lower burners 41
in service. Consequently, the high temperature superheaters and reheaters gain
more heat which raises the steam temperature.
Burner tilt angle. The angle at which the fuel and air is introduced into the furnace
affects the position of the fireball, the furnace exit temperature, and consequently,
steam temperatures.
Excess air. Changing excess air quantity affects steam temperature, due to the
influence of gas velocity on convective heat transfer and also due to the cooling
effect on the furnace temperature.
Feed water temperature. Superheat increases with a decrease in feed water
temperature. For a given firing rate, a decrease in feed water temperature reduces
the quantity of steam produced. The increased amount of heat discharged per unit
of steam raises the superheat. The removal of feed water heaters from service for
maintenance has the most severe effect on feed water temperature [28].
Blowdown. Removal of heat by means of blowdown increases the firing rate per
unit of steam produced and therefore increases the steam temperature. The effect
here is the same as a decrease in feed water temperature [28].
Steam bleed. The use of saturated steam or steam with low superheat for
auxiliaries increases. the firing rate per unit of steam after the bleed point and
therefore increases the steam temperatute.
3.2.2 Long time lags
The speed of control response depends on the amount of dead time and time lag in a
system [41]. A quantitative method of expressing the speed of control is to take the
integral over time of the absolute error in controlled variable after a disturbance. The
measure, called Integral of Absolute Error, or IAE, is a representation of the extent of an
excursion from setpoint combined with its duration. For systems comprising both dead
time and time lag, Shinskey [42] shows that the theoretical lowest IAE is:
(3.1) MEmin = 1Kp Aql 'EP - e 1') where: rd = dead time
r, = time constant, i.e.
42
(time for temperature to reach 0.632 of its final value) - rd Kp = process gain, i.e. relation of temperature to disturbance tlq = magnitude of disturbance
For example, measurements made on a reheater of a 686 MW boiler at full load have indicated a process dead time of 2 minutes and a time constant of 5.5 minutes between a change in desuperheat flow rate and reheater outlet temperature (Figure 3.4). Valsalam [43] documented process lags (time lag + dead time) of 8 - 10 minutes.
60
40 ID
0
20
0 I I I -5 0 5 10 15 20 25 30 35 Minutes
Spray valve position [%] Deshtr out temp [300 - 400 deg C]
— Reheater out temp [450 - 550 deg C]
Figure 3.4 Reheater outlet temperature reacting to increased spray water flow.
The minimum IAE for the reheater recovering after a disturbance causing a 10°C deviation is:
IAEppp(10) = 10 * 2.2 (1 - e 5.5) (3.2) = 6.7 min °C (3.3) 43
The real attainable IAE may be significantly more than the theoretical minimum, depending on how the controller is set up [42].
The reason for the slowness of the process lies in the thermal inertia mechanically present in the plant. Consider a reheater of a 686 MW power plant with an internal volume of 325 m3, an average working pressure of 4 MPa, an inlet temperature of 300°C, an outlet temperature of 540°C, and a steam flow rate of 500 kg/s. The following details regarding the reheater applies [26], [46]:
Mass of steam inside reheater tubing & headers: Mtteam = 4 276 kg Average steam temperature (assumed): Tiftaff, = 420°C
Specific heat of steam at 4 MPa & 420°C: Cp steam 2.314 kJ/kg°C
Total heat capacity of the steam: steam * cp.,„a„, = 9.9 MJPC
Mass of reheater tubing and header material: M„,„„,= 540 000 kg
Specific heat of 1.5 % carbon steel: Cp metal . 0.486 kJ/kg°C
Al Total heat capacity of the steel: * Cp metal = 262 MJ/°C
Therefore, it requires 9.9 MJ heat to raise thelemperature of the steam by one degree Celsius, while a similar rise in temperature for the steel requires 262 MJ heat. This implies that, during the process of correcting a reheat steam temperature deviation, 96.4 % of the control action is absorbed by the reheater metal, while 3.6 % of thedontrol action effectively changes the steam temperature.
3.2.3 Process parameters varying with time The process parameter that varies most significantly with time is that of heat resistance due to the ash and slag deposits on the heat transfer surfaces. The rate of contamination depends on the ash content of the fuel burnt, ash properties, boiler load and furnace temperature. High pressure steam is utilized to clean the surface of boiler components so that heat transfer is improved. Cleaning evaporator surfaces ahead of the superheater will reduce the gas temperature and produce more steam. This will tend to decrease the degree 44 of superheat. Cleaning superheater surfaces will increase superheater absorption and raise steam temperature.
3.2.4 Process parameters varying with load Three major load based nonlinearities affect temperature control in large power boilers: process time constants, heat transfer and heat absorption.
Process time constants The first load based nonlinearity affecting temperature control, the process time constant, shortens as the boiler load increases. This reduction in system lag is due to the increased steam flow rate which carries changes in steam temperature through to the superheater and reheater outlets faster at higher boiler loads. Figure 3.5 shows the effect of a decrease of desuperheating on the Kendal reheater at two load points. The system response is visibly faster at the higher load point. Note that the difference in final temperatures results from different changes in desuperheating during the two tests. Valsalam [43] also describes large changes in process parameters based on changes in load.
70
0 co co v 60 rn Lc) U) "6 50
e) 0.
40 I I I -5 0 5 10 15 20 25 30 35 Minutes
100 % boiler load — 70 % boiler load
Figure 3.5 Reheater outlet temperature response under two load conditions. 45
Heat transfer rate As discussed previously, convective heat transfer increases in relation to boiler load while radiant heat transfer decreases. This nonlinearity may be cancelled by balancing convective and radiant heat transfer surfaces during the boiler design. However, should this balance be suboptimal in practice, the effect on steam temperatures can be significant and small variations in boiler load may place great strain on the steam temperature control system.
Steam properties The volume of steam in a superheater or reheater remains virtually constant through boiler loads. This is however not so for the mass of steam in these components, as the density. of steam changes with pressure. At higher boiler pressures, the increased mass of steam in the superheater and reheater will reduce the effect of short term heat transfer variations on steam temperature. Also, at higher loads, the increase in steam flow will require an increase in desuperheating spray water flow to achieve the same control action. The steam temperature controller gains should therefore be adjustable on-line to achieve consistent control results. Steam properties also change through pressure. At 10 MPa, spray water requires 1872.14 kJ/kg to boil from 200°C while at 16 MPa it requires 1726.28 kJ/kg - about 10% less energy. This change in energy requirement results in an increase in spray water needed for desuperheating at higher steam pressures.
3.2.5 Control loop interaction Many control loops exist on a large power boiler. The generator load controller manipulates steam flow to the turbine via governor valves to control generator power output. Boiler pressure is controlled by manipulating the fuel firing rate and steam temperature is controlled via one or more of the various methods discussed previously [44].
A high degree of interaction exists among the control loops, and in most cases, steam is the common denominator [40]. For example, a reduction in steam flow rate from the load controller will result in an increasing boiler pressure and increasing steam temperatures. 46
Steam temperature control via desuperheating increases steam flow. A reduction in fuel firing rate also reduces steam temperatures (at least in the short term). It is therefore not possible to control just one variable while disregarding the others. Similar interactions between process systems are also described in [45]. The result of interaction is that two or more controllers may start cycling continuously, because of phase differences in their control objectives. A good example here is the following description of cycling caused by reheater temperature control.
Assume the reheater outlet temperature is above its setpoint. Desuperheating spray water is injected into the reheater to reduce the temperature and bring it to setpoint. The added mass of water increases the steam flow rate to the intermediate and low pressure turbines, which increases generator load. The load controller responds by closing the governor valves which reduces the steam flow rate. Asa result of the reduction in steam flow rate, the boiler pressure increases and consequently, the steam pressure controller reduces the fuel firing rate. The reduction in fuel flow rate decreases the reheater heat pickup and the outlet temperature decreases below its setpoint. The entire cycle is repeated in the reverse and may continue cycling or even become unstable unless an operator intervenes manually. A similar description of system interaction is also given in [8].
3.2.6 Over-firing When a power generating unit needs to move from one load point to the next, the fuel flow rate needs to be manipulated to effect the load change. Due to the thermal inertia of the boiler; the change in steam flow rate will lag behind the change iniuel flow rate. To make the generator load follow a predetermined load ramp rate, steam flow must be increased proportionally. To overcome the time lags inherent in the boiler, it is necessary to inject a substantial quantity of additional fuel during the initial stages of the load ramp. This technique is called over-firing, and the magnitude of over-firing is dependent on the load ramp rate. In the case of the Kendal boilers, a 5% per minute load ramp rate requires almost 20% over-firing [36].
The additional heat injected into the boiler is used to overcome the thermal inertia of the 47
steel pipework of the boiler and the water or steam that flows through it. At this point, the concept of relative thermal inertia will be introduced as the ratio between thermal inertia and heat transfer:
thermal inertia Ir - (3.4) heat transfer
If the Ir of all the different boiler components are not equal, the components with the lower Ir will react more severely to over-firing than those with a higher Ir. For example, if the Jr of the evaporator is greater than that of the superheateror reheater, changes in steam production during transients will be lower than changes in heat transfer to the superheating elements. During a load increase, where over-firing is a positive quantity; too little steam will be produced for sufficient cooling through the superheater and reheater. Consequently, the final steam temperatures will be raised and this increases the burden on the steam temperature controllers.
3.3 Temperature excursion study A study was done at Kendal to establish the extent of the problem with steam temperature excursions and their causes. During February, March, and April 1996, superheater and reheater steam temperatures were monitored over a period of 213 unit-days and all temperature excursions were recorded and compiled into a list. The study showed that on average, 1.81 steam temperature excursions occurred per unit per day. Table 3.2 summarises the findings of the study.
Criterion Number of excursions Main steam temperature > 551 °C 71 Main steam temperature gradient > High 252 Hot reheat temperature > 555 °C 61 Hot reheat temperature > 565 °C 2 Total number of excursions 386
Table 3.2 Results of excursion study. 48
Of the 386 temperature excursions, 220 had operator-logged explanations as to why the particular excursion occurred. The two most common reasons for the excursions were mill changes / trips and unit load ramps. The remainder of the excursions were due to instrumentation faults, unit start-ups or shut-downs, special tests, wet coal, capability load runbacks, etc. Figure 3.6 gives a Pie-chart representation of the weighting of the different causes of temperature excursions at Kendal.
The two main causes of temperature excursions, load ramps and mill changes, will be discussed in more detail below.
Figure 3.6 Causes of temperature excursions at Kendal.
3.3.1 Load ramps
The loading rate of the units at Kendal is adjustable between 0 and 35 MW/min, but is is normally set to 15 MW/min for load changes. The magnitude of the load ramp is determined by the national dispatch centre, based on customer demands and the location and size of power station units on the power grid. A power generating unit may undergo hundreds of load changes daily, ranging between 10MW and 100MW in magnitude.
These load variations have a significant disturbance on steam temperature (Figure 3.7).
When generated load needs to be altered, the boiler fuel flow must be adjusted first. The 49
firing rate of all mills are adjusted simultaneously, therefore, the heat transfer to all boiler
components are changed simultaneously. However, the steam flow rate lags behind the
increased firing rate, due to the boiler's thermal inertia. This results in a change in the
heat transfer to the superheater and reheater before the steam flow changes, leading to overheating or cooling of the steam.
An example of the effects of a load ramp is demonstrated by the following three tables.
The data was taken from an actual unit capability test of an up ramp in load from 60% load (412 MW) to 80% load (549 MW) at a loading rate of 20 MW / min. Heat transfer
rates were obtained by means of a neural network heat transfer model (Appendix C).
' 3 LH SHTR OUT TrIP RH SHTR OUT. TMP
GEN MW
Figure 3.7 Main steam temperature deviations from setpoint caused by load variations. Recorded over the 24 hours of 1996.06.28.
The steady state conditions before the ramp commenced, are shown in Table 3.3. 50
Furnace conditions Heat transfer A-mill demand 48% Evaporator 480 MJ/s B-mill demand 48% Superheater 315 MJ/s C-mill demand 48% Reheater 152 MJ/s D-mill demand 50% Main steam flow 294 kg/s E-mill demand 0% Total fuel flow 60 % 02 setpoint 4% Burner tilt angle -15°
Table 3.3 Steady state conditions before the ramp.
About 5 minutes into the load ramp, the conditions have changed drastically from what they were before the ramp. Apart from the increased fuel flow and steam flow, the burner tilt angle was decreased via automatic control, due to high temperatures on the superheater, and the excess air was reduced via automatic control due to high temperatures on the reheater. The results are shown in Table 3.4.
Furnace conditions Heat transfer A-mill demand 63% Evaporator 604 MJ/s B-mill demand 63% Superheater 411 MJ/s C-mill demand 63% Reheater 214 MJ/s D-mill demand 63% Main steam flow 1351 kg/s E-mill demand 0% Total fuel flow 80 % 02 setpoint 3% Burner tilt angle -19°
Table 3.4 Conditions during ramp.
During the load ramp, large disturbances in equilibrium are caused due to changes in heat transfer without similar changes in steam flow. The heat imbalance during a load ramp is illustrated in Table 3.5. Five minutes after starting the load ramp the fuel flow had increased by 33% while the steam flow had only increased by 19%. 51
Boiler element Before ramp During ramp Difference Total fuel flow 60% 80% 33% Main Steam flow 294 351 19% Evaporator 480 MJ/s 604 MJ/s 26% Superheater 315 MJ/s 411 MJ/s 30% Reheater 152 MJ/s 214 MJ/s 41% Table 3.5 Changes in heat transfer during load ramp.
Although some over-firing is needed to overcome the thermal inertia of the boiler mass, the over-firing should be proportional to the mass-related thermal inertia of the different boiler elements. The evaporator has more than double the mass of the superheater or reheater [46], but the heat transfer to it during load ramps is far below double (Table 3.5). This maldistribution of heat leads to superheater and reheater steam outlet temperature deviations from setpoint.
Strong feedforward signals based on load gradient is used to bias the desuperheating on superheater and reheater. These feedforwards have been tuned to counteract most of the effect that over-firing has on Steam temperatures, but even with the feedforwards, typical temperature excursions during 15 MW/min load ramps are 8°C on the superheater and 13°C on the reheater. The actual capability test described considered here, had an increase of 12°C on the superheater and 17°C on the reheater.
Asa result of these temperature excursions, the load ramp rate of the Kendal units have been restricted to 15 MW/min as opposed to the contractual specification of 35MW/min. With the reduced load ramp rate, it was possible to maintain the superheater outlet temperatures within the specified ± 11,2°C from setpoint. However, the reheater outlet temperature still exceeded the originally specified ± 11,2°C margin, but the contractual specification was since relaxed to ± 17°C for the reheater only.
3.3.2 Mill changes / trips The other major contributor to steam temperature excursions is coal mill changes. When 52 a mill is shut down or started up, the fireball in the furnace is shifted or distorted because the fuel injection points have changed. The shifting of the fireball changes the furnace-to- boiler heat transfer pattern. A shift in the heat transfer pattern may increase or decrease the heat transfer to the superheater and reheater, depending on the change in heat transfer to these components, thereby affecting the steam temperature.
When a mill is taken out of service or trips, its fuel flow decreases to zero, while the total boiler fuel demand remains virtually unchanged. Therefore, the fuel demanded from the mills remaining in service is changed proportionally to compensate for the loss of fuel from the tripped mill. The opposite is true for placing a mill in service.
M . ; 1 es
LH S H T R '1'.0.11T • IMP RH: -$..HIR:10.0T: .
Figure 3.8 Temperature excursion caused by a mill shut down. Recorded over one hour.
The transfer of heat from the furnace to the boiler components is sensitive to mill combination and relative mill loading because the fuel from each mill is injected at a different elevation in the furnace. Therefore, an upset in heat distribution accompanies a 53
mill change or mill trip. This leads to a disturbance in the equilibrium in heat transfer
needed for maintaining stable steam temperatures. The steam temperature changes due
to the disturbance (Figure 3.8) and the control system responds by injecting more or less
desuperheating spray water. The steam temperature control system cannot anticipate the disturbance in heat distribution and has to wait for the steam temperature to change before it can respond.
A typical example is a mill trip on a Kendal unit running at 586 MW. Before the trip, four
mills are in service, say A, B, D, and E. The 0 2 content in the flue gas is 3% and the
burner tilts are angled at 0°. The heat transfer to the boiler elements under these
conditions were obtained via a neural network heat distribution model (Appendix C) and are shown in Table 3.6.
Furnace conditions Heat transfer
A-mill demand 70% Evaporator 662 MJ/s
B-mill demand 70% Superheater 463 MJ/s
C-mill demand 0% Reheater 242 MJ/s
D-mill demand 70%
E-mill demand 70%
02 setpoint 3%
Burner tilt angle 0°
Table 3.6 Conditions before mill trip
Now consider a trip of the D-mill. Its fuel flow decreases to zero while the other mills all increase production from 70% to 93.3% to absorb the deficit in total fuel flow. The burner tilts adjust automatically to -15° to compensate for the higher average position of fuel injection. At first, the 02 concentration will remain unchanged (apart from transients) but it will start reacting slowly on changes in reheater spray water flow rate. Table 3.7 shows the new furnace conditions and the resultant heat transfers shortly after the mill trip. 54
Furnace conditions Heat transfer
A-mill demand 93.3% Evaporator 656 MJ/s
B-mill demand 93.3% Superheater 511 MJ/s
C-mill demand 0% Reheater 200 MJ/s
D-mill demand 0%
E-mill demand 93.3%
02 setpoint 3%
Burner tilt angle -15°
Table 3.7 Conditions after mill trip
Due to the mill trip, the disturbance in furnace conditions has a major effect on heat pickup in the boiler. Heat transfer to the evaporator decreased, heat transfer to the superheater increased, and heat transfer to the reheater decreased. The variations in heat transfer are summarized in Table 3.8.
Boiler element Pre-trip Heat Tx Post-trip Heat Tx Delta Heat Tx
Evaporator , 662 MJ/s 656 MJ/s - 6 MJ/s
Superheater 463 MJ/s 511 MJ/s + 48 MJ/s
Reheater 242 MJ/s 200 MJ/s - 42 MJ/s
Table 3.8 Changes in heat transfer caused by a mill trip.
These changes in heat transfer to the different boiler elements cause temperature excursions. Capability tests have shown steam temperatures to change by as much as
20°C after mill trips. Temperature excursions are also caused by normal mill shut downs, as shown in Figure 3.8.
There is no way for /he installed control system to act directly on changes in heat transfer because it is not measured. The only method of automatic compensation used, is by waiting for changes in temperature and then adjusting the degree of desuperheating accordingly. Temperature excursions as a result of coal mill disturbances are also 55
reported by Aitchison e.a. [39] & Franchot [47].
3.4 Instrumentation and control configuration Irrespective of the type of control algorithm used: PID or advanced, the control strategies are all based on process measurements, control calculations, and control actions. This section describes the instrumentation and control element configurations used for implementing steam temperature control.
3.4.1 Basic closed loop control The most basic arrangement for steam temperature control is as follows: Measure the steam temperature at the point of exit from the boiler. Compare the temperature measurement to the steam temperature setpoint. Use the error between setpoint and measurement to calculate a control action. Drive the physical control element according to the desired control action.
To turbine A Temperature Setpoint
Tern perature Superheaters Measurement
Controller Adjustment to control element \ /
\ / Figure 3.9 Basic temperature control loop.
The control element may be the desuperheater spray flow control valve, burner tilt positioner, bypass darhper, etc, depending on the boiler design. This basic control setup, referred to by the ISMC [38] as single element control is shown in Figure 3.9. It is 56 recommended that single element control only be used as the sole method of control in applications with slow load changes, or where steam temperature is not critical [38].
3.4.2 Cascade control
When the output of one controller is used to drive the setpoint of another, the controllers are said to be cascaded [48]. If desuperheating is used for steam temperature control, a cascade configuration is recommended to reduce the system nonlinearities and improve its disturbance rejection capabilities [8], [38]. Cascade control is recommended where:
spray water is the primary method of steam temperature control;
variable steam pressures exist;
the spray water supply pressure may vary; and
the spray water control valve has a nonlinear characteristic.
The outer loop (or master) controller compares the steam temperature to the setpoint and its output drives the setpoint to the inner loop (or slave) controller. The slave controller measures desuperheater outlet temperature or spray water flow rate, compares it to the setpoint received from the master controller and drives the desuperheater spray water control valve.
Having the inner loop control spray water flow, results in a system immune to changes in spray water pressure. Having the inner loop control desuperheater outlet temperature makes the system immune to changes in both spray water pressure and steam flow rate.
Therefore, the preferred method is to control desuperheater outlet temperature [38]. The arrangement is shown in Figure 3.10. 57
To turbine Steam Desuperheater tern perature setpoint
Main steam temperature
y y measurement Master Desuperheater controller outlet temperature measurement Desuperheater temperature Spray water control valve 'setpoint Y Y Slave controller Adjustment to control valve
Figure 3.10 Cascade control arrangement.
3.4.3 Feedforward control
A powerful method used for disturbance rejection is the of feedforward control [48]. . In its simplest form, feedforward control measures a disturbance, calculates the magnitude of control action needed to counteract the disturbance, and sends this magnitude as a bias to the control element (or to the slave controller in the case of cascade control). For power plant steam temperature control, the main feedforward signal may be derived from the boiler load index, but it is recommended that the feedforward be based on all major influences on steam temperature, including adjustments to heat distribution within the boiler and changes in the thermodynamic properties of steam [38].
Many advanced control strategies applied to power plant boilers use a disturbance calculation for a feedforward signal to cancel out the effect of rapid load changes on steam temperature. This concept, outlined in Figure 3.11, will be discussed in more detail later and it will be shown that feedforward signals can be used to control the entire heat distribution pattern of a power plant furnace. 58
To turbine
t Auxiliary process Boiler measurement(s)
Y Y L
Feedforward Adjustment to control control element(s) \ / calculation(s
Figure 3.11 Feedforward control.
3.4.4 Combination of control configurations Feedforward control is not used as the sole means of temperature control because it measures process disturbances rather than steam temperature itself. If an unmeasured disturbance occurs (for instance the sooting of boiler tubes), the steam temperature deviation will not be corrected because feedforward control does not take steam temperature measurements into account. In other to control steam temperature in the reality of unmeasured disturbances and nonlinearities, feedforward control is combined with feedback control.
The combination of feedforward and feedback control is done by adding the feedforward signal to the output of the steam temperature controller so that both control modes have access to the final control element used for steam temperature control [8]. In the case of cascade control being used, the sum of the feedforward and feedback control modes forms the setpoint to the slave controller (Figure 3.12).
59
Steam Main steam To turbine tern perature temperature setpoint measurement
Desuperheater YY Steam temp controller
Desuperheater Feedforward outlet controller temperature measurement Spray water control valve
Cascade controller Adjustment to Desuperheater)111 control valve tem perature setpoint
Figure 3.12 Combined feedback, feedforward and cascade control arrangement.
3.4.5 Multiple control elements It is possible to use more than one steam temperature control device on a single superheater or reheater. The individual control loops can be configured to operate independent of each other, the control action can be distributed between control elements, or individual control loops may operate in a coupled fashion [38].
Independent Control Where more than one control element is used for steam temperature control and the configuration of the controllers is to operate independently of each other, then each has a different control objective. For example, on a superheater with two-stage desuperheating, the first stage desuperheater may be used to control the outlet temperature of an intermediate stage of the superheater while the second stage desuperheater will control the final steam temperature [38]. Although the second stage desuperheater is influenced by actions of the first, the two control loops run independent of each other. Feedforward control signals should be passed only to the control element responsible for controlling the final outlet steam temperature [38]. 60
Distributed Control The control action may be distributed among the available control elements. An example of distributed control would be where the spray water flow rate is divided between the first-stage and second-stage desuperheaters on a superheater [38]. A single steam temperature control loop exists but its control action is divided and passed on to two cascade secondary controllers. The primary control action is divided according to a ratio recommended by the manufacturer. Feedforward signals are added to the output of the steam temperature master controller and is thus distributed between the cascade slave controllers in the same ratio'as the closed loop control signal [38].
Coupled Control In the coupled control strategy, two closed loop controllers are linked (not cascaded). The control strategy is divided into a fast primary action and a slower secondary action. For example, if desuperheating and burner tilt angle is used to control reheater temperature, the desuperheating is used as the fast primary control method and burner tilt angle provides slower secondary control [42].
Under these conditions the primary controller is used for steam temperature control. It receives steam temperature feedback from the plant, compares this to its setpoint and drives the desuperheater control valve, either directly or through a cascade arrangement. However, the burner tilt controller is assigned a setpoint representing:some optimal amount of spray water flow needed for good control. The tilt controller compares this setpoint to the actual spray water flow rate and adjusts the burner tilt position according to the error between the two. Changing the burner tilt angle will affect the steam temperature and consequently the spray water flow rate.
The same arrangement is used when two sets of desuperheaters are installed in series between different stages of a superheater. Figure 3.13 shows the arrangement where more than one final control element is used to adjust the same steam temperature. Feedforward signals are passed on to the primary (fast) controller only. 61
To turbine Primary control element setpoint
Tern perature Setpoint 4 Tem perature Measurement Superheaters
Prim ary controller Adjustment to primary control element
y Secondary Adjustment to controller . secondary control / elem ent Figure 3.13 Multiple control elements with coupled control.
3.4.6 Saturation protection Regardless of which control structure is used, some means of control protection should be provided to prevent the outlet conditions of a desuperheater from reaching saturation [38]. This is mainly to protect the turbine from receiving wet steam from the boiler, and to prevent the scaling or erosion of the inner walls of the superheater tubes. The limit on desuperheater outlet temperature may even be set marginally (i.e. 10 °C) above the saturation temperature [44].
3.5 Developments in steam temperature control The function of the steam temperature control system is to maintain the temperature of the steam within the boiler or turbine manufacturer's specified limits. Generally the goal is to obtain a specified final steam temperature over the entire boiler load range, but there may be deviations from the rule. For instance, the steam temperature setpoint may be decreased at very low boiler loads in large sized boilers [4].
Although modem control system design methods can improve the dynamic behaviour of many processes, classical PM controllers are still most widely used [49]. This is true even despite the increased programming capabilities of modem digital control hardware - allowing the
62
implementation of complex control strategies. Peters [49] motivates two probable reasons for this: Many advanced control strategies require much time and special skills to design implement and optimize.
The difficulty to understand complex control techniques promotes a lack of interest in advanced control systems - especially among plant personnel.
Although HD controllers still outnumber the advanced control installations, requirements for
increased load manoeuverability, reduced emission levels, and increased cost-efficiency is
demanding control capabilities only possible with advanted control strategies [50]. Studies on
the application of advanced control to the process industries show savings of between 2% and 6%
of annual operating costs [51]. The (documented) advanced steam temperature control schemes
recently tested or Stalled on power boilers are all in some way either nonlinear / adaptive, model-
based / predictive, or both. The reasons for these trends are given below:
a) Nonlinear / adaptive control
The change in process characteristics between different loading points gives rise to different
control parameters needed for good control. For instance, less spray water flow is required to
correct a temperature deviation at low steam flows compared to correcting the same deviation
at high steam flows. A controller needs to be either nonlinear or it needs to adapt itself to the
changing process characteristics to provide optimal control throughout the operating range.
• b) Model based / predictive control
The controller needs to know beforehand how a process will react to a disturbance lso that it can counteract this disturbance before its effects become apparent. For example, because increased
air flow leads to increased convective heat transfer, a model based / predictive controller will
automatically increase desuperheating after an increase in air flow.
Some of the tested and documented advanced control techniques for steam temperature control
on power plant boilers are described below. 63
3.5.1 Advanced PID control
It was already said that PID is still used more than any form of advanced control. Even
so, with the existing structure of a PID controller, advances are being made in this field
by gain scheduling on the PID controller or in obtaining optimum controller settings.
Gain scheduling
A HD controller is a linear controller and is therefore not well suited to controlling plants
with major nonlinear characteristics or changing parameters [42]. However, modern
microprocessor-based control systems have made it possible to adjust the P, I, and D
settings of a controller automatically, and on-line. The adjustment of the controller
settings are done based on the measurement of some index variable through which the
plant is nonlinear. In many cases, only the controller gain needs to be adjusted [48]. The
advantage of gain scheduling is that the adjustment in controller settings on an essentially
linear PID controller ensures more comparable control actions throughout a nonlinear operating envelope.
Improvements to steam temperature regulation by using gain scheduling are commented
on by Hitz e.a. [54]. Tests were run on a simulation of an Eraring boiler and HD control
with gain scheduling showed a marked improvement over fixed gain control with respect to steam temperature regulation.
Optimal PID control
The development of optimal control techniques for a standard Cascade steam temperature control arrangement is described by Peters e.a. [49]. Firstly, a nonlinear model of the process and controller was created and verified against the real system. The model was linearized around the 90% load point and converted to a state-space model. A quadratic criterion function was then minimized with the controller parameters as variables and optimal controller settings were obtained.
3.5.2 Model-based predictive control
The control algorithm is based on the use of an on-line boiler model as a predictor of 64
boiler behaviour. At each time step the plant inputs are sampled and used as inputs to the model. The model runs through a number of time steps in order to produce a prediction of boiler states and steam temperature ahead of time. The predicted value of steam
temperature is compared to the setpoint and the error is used for control.
This technique was used on a 217 MW nuclear reactor for steam temperature control
[52]. The results showed temperature deviations of less than 3°C during on-line refuelling
transients. Although no indication of the degree of improvement in temperature
regulation is provided, the control scheme was rated so successful that it was also installed on other identical units.
3.5.3 Advanced feedforwards
Feedforward control is in essence a model-based control. By incorporating a process
model the steam temperature control system is able to anticipate future changes in steam
temperature associated with current changes in furnace or boiler conditions. The controls
can then take corrective action before an upset in the steam temperature actually occurs. In this way, negative effects of the process deadtime can be reduced.
Aitchison e. a. [39] documented how attempts at a dynamic feedforward using steam flow
and air flow to modify the desuperheater outlet temperature setpoint during transient
operation proved to be unsuccessful. The reason given was difficulties in establishing the
dynamic process models required for accurately tuning the feedforward signals. It was difficult to obtain repeatable test data using simple step response type tests.
A more heuristic approach based on the analysis of the system physics was also documented by Aitchison e.a. [39]. The feedforward was based on a dynamic enthalpy calculation to determine the required secondary superheater inlet temperature setpoint for various pressure and temperature operating conditions. The secondary superheater inlet enthalpy requirements were calculated as the main steam enthalpy less the enthalpy rise across the secondary superheater. The model of expected enthalpy rise received steam flow, gas recirculation flow, air flow, and main steam pressure as its inputs and were 65 calibrated at three load points to account for changing process parameters.
A similar approach was followed by Hitz e.a. [53] and major improvements over PliD control were observed on a boiler simulation. The integral of absolute error was reduced by a factor of 3 to 5 across ten test scenarios.
Zhu e.a. [45] document the success achieved with a model reference feedforward controller installed on the 565 MW Unit 1 of Virginia Power's . Mt. Storm power station.
The loading rate of this plant could be doubled and the final steam temperature setpoint could be increased by 2.8°C• due to reduced temperature overshoot. The new control strategy also reduced the need for manual spray action during severe load transients.
Also documented by Zhu e.a. [45] are results obtained with the same control strategy on a boiler simulation of Virginia Power's Chesterfield power station. This unit has desuperheaters for controlling main steam temperature and tiltable burner nozzles for controlling reheat steam temperature. Simulation results showed a potential improvement in control of superheat temperatures that markedly reduced the amplitude of variations in steam temperature and fuel control loops.
The successffil implementation of an advanced feedforward control system on the 660MW units at Eraring power station in Australia is documented by Hitz e.a. [54]. During load changes, the steam temperature controller output is supplemented by a feedforward signal based on a heat balance calculation for each superheater. The feedforward inns to control steam temperature by balancing the heat flow into the superheater with the heat carried away by the steam. Heat flow into the superheater is inferred from boiler air flow, as the fuel flow signal at Eraring is very noisy. The heat uptake of the steam in each of the three superheaters is continuously computed from the inlet and outlet steam conditions and spray water conditions. The heat transfer of each superheater is passed through a low- pass filter to obtain a moving average. Boiler air flow is similarly filtered to obtain a moving average. 66
The ratios of heat transfer to air flow for each superheater are initialized with values determined experimentally and are updated on-line by comparing conditions before and after a ramp in load. Thus the ratios used in the current transient are determined with data from the previous transient. This scheme is made feasible by the mode of operation of the Eraring units, in which substantial load changes are usually followed by periods of steady- state operation. Because the quantity of spray water needed to control the steam temperature during load transients is computed from enthalpy and a heat balance, the system automatically compensates for changes in cooling capacity of the spray water (with steam flow, temperature and pressure and spray water temperature), thereby accounting for the effect of nonlinearities.
A prototype of the control system has been in operation in Eraring since January 1991. Unit loading rates are still restricted to 7 MW/min at loads between 200 and 350 MW, but superheater outlet temperature is now controlled within ±2°C from setpoint. The loading rate restriction is due to spray system saturation at low loads. At boiler loads above 400MW, load ramps of 100MW at a rate of 20MW/rnin can be performed while controlling the steam temperature within ±2°C from setpoint. This is a vast improvement over the previous ±12°C deviation from setpoint.
Another successful installation of model reference feedforward control is documented by Franchot e.a. [47]. This controller was installed on Canal Electric Company's 600 MW Unit No. 2. It consists of a PI controller supplemented by a feedforward signal generated by a nonlinear mathematical model to predict the required desuperheater spray flow rate. The PI controller ensures robustness during changes in process constants that were not accounted for.
The model considers combustion, furnace performance and heat transfer and was derived using first principles. Most model parameters were determined using equipment design characteristics, but certain model constants were determined empirically during on-line model calibration. 67
The advanced control strategy implemented on Canal Unit 2 decreased temperature
excursions of between 28°C and 55°C under normal control to below 14°C. The range
limits on automatic (remote) loading could also be increased from 200 - 530MW to 60 - 600MW.
3.5.4 Optimal control
Optimal control is a model-based control strategy - a model of the process is used to
calculate the optimal control action [39]. The process model consists of a set of state
equations and may be derived by means of analysis and differential equations or by means
of plant measurements and statistics [40].
Due to the high complexity and low accuracy of the analytical method, state equations are
normally obtained by means of tests performed on actual plant [39],[40]. This is done by
injecting pseudo random binary test signals into the system using each of the manipulated
variables in turn. A computer logs the manipulated variables and the state variables, and
uses multivariable autoregression to fit the data to a mathematical model of the plant
dynamics. The order of the model is chosen to minimize the regression error [40]. The
mathematical model is then transformed into a state equation. Once state equation is
defined, the optimal state feedback gain matrix is determined. For this determination, a
digital simulation technique is utilized in which the state equation and a candidate gain
matrix are used at each control interval. The method uses dynamic programming for
adjusting the gain matrix to minimize a quadratic criterion function.
Changes in plant characteristics due to boiler sooting, or nonlinearities based on operating
point, can lead to a deterioration in the performance of an optimal controller [55]. This
happens because of the discrepancy between the behaviour of the actual plant and that of
the state-space mathematical expression of the plant dynamics. The problem of
nonlinearity can be overcome by creating different process models, valid at different
operating points, and interpolating between the parameters of these models, based on the current operating point [39],[40]. Plant characteristics that change due to unmeasured disturbances (e.g. sooting of boiler tubes), remains a problem for pure (non-adaptive) 68
optimal control. The addition of integral control action on controlled variables can
compensate for these model mismatches and eliminate steady-state errors [56].
An optimal regulator was first implemented on a thermal power plant in February 1978 at Buzen Power Station No. 1, a 500 MW plant of Kyushu Electric Company of Japan
[40]. As of 1987, Kyushu has optimal control in operation on five power plants. The
improvement in performance realized by the optimal regulator was quite significant [40].
With optimal control the size of load ramps could be doubled and still the steam temperature deviations remained less than half of that obtained under P133 control.
The implementation of, and results obtained with an optimal controller for steam temperature regulation, called the ACORD system, is described by Aitchison [39]. The first installation of ACORD was at the Sendai plant in Japan, and the second was completed during February 1991 at the 500 MW Babcock & Wilcox boiler of Ontario
Hydro's Nanticoke Unit 7. With the ACORD system, a significant reduction in temperature deviation was achieved while the maximum ramp rate had been increased simultaneously. With conventional controls, the maximum boiler pressure ramp rate of
150kPahnin resulted in temperature deviations of -20°C to +11°C. With ACORD on, the increased ramp rate of 200kPaimin resulted in temperature deviations of only --9°C to +9°C.
Several optimal controllers can be configured to operate in parallel, each controller for a specific process variable. Hanson e.a. [57] describe the design, installation'and testing of an array of four optimal controllers, controlling left-hand and right-hand superheat steam temperatures, reheat steam temperature, and furnace gas outlet temperature (the latter is used to control NO„ emissions). Every controller receives inputs from the other controllers' outputs, effectively decoupling interaction between the different control loops.
Hanson e.a. [57] also report that after the installation of this advanced control strategy in
1994 on Montana-Dakota Utilities' 45 MW Coyote power station, encouraging improvements in steam temperature control and general control system stability was evident. Previous temperature swings of ±11°C under steady state conditions and ±22°C 69 during transients were reduced to ±1.7°C and ±3.9°C, respectively.
3.5.5 State variable control with observer The concept of state variable control is to use additional state measurements from the process (e.g., intermediate temperatures along the superheater or reheater) to provide more accurate control action [58]. The necessary state measurements are not available on a power plant, but they can be simulated by means of a dynamic process model - called an observer [59], or a Luenberger observer [60]. The arrangement is referred to as State variable Control with Observer (SCO).
The observer consists of a series of first order lags [53]. Each of the lags has an • associated gain and a time constant. The outputs of the lags simulate various temperatures along the superheater. The simulated temperature signals are multiplied by individual gain factors and summed to create a control feedback signal. SCO control provides proportional action only. Therefore, a steady-state error will exist unless integral control is used to trim the control action [53].
Practical results obtained with SCO on a reheater at Kendal power station, showed an average improvement of 20% to 50% over PID control in reheater temperature deviations during load ramps. SCO control was also extensively tested on a nonlinear simulation of a 250MW unit at Cromby Power Station of Philadelphia Electric Company. Only minor improvements over PI control was observed [53].
3.5.6 Adaptive control An adaptive controller continuously monitors the characteristics of the plant it controls and automatically adjusts its controller parameters to maintain some predefined performance. In this way, an adaptive controller can adapt its control actions according to variations in plant parameters. Due to the changes in power plant characteristics through time and operating point, the adaptive controller would seem well suited to achieve the desired control action regardless of plant dynamics [55]. 70
Adaptive control essentially consists of three parts: a state observer, an adaptive plant identifier, and a controller with adaptable parameters, but it may have a fourth part, a model of the desired plant response [61]. The observer is a state variable filter, for extracting the plant state information. The identifier determines the parameters inside a predefined transfer function of the plant from this state information and the error between the actual plant and the estimation [62]. The desired response model may be a criterion for stability [43], or it could be a transfer function containing some predefined plant response to a setpoint change. The adaptive controller parameters are adjusted according to the desired response and the transfer function of the identified plant. It is necessary to know the order of the plant beforehand as the order of the plant determines the order of the reference model and the order of the controller.
It has already been shown how adaptive control is superior to PM control in the nonlinear application of power plant drum level control [63]. Simulation studies done by Nomura e.a. [64] showed that with adaptive control the deviations of steam temperatures from setpoint could be reduced to half of that obtained with conventional PlD control. The design, application and testing of an adaptive steam temperature controller on two different 375 MW power plants (Nishi-Nagoya and Owase-Mita) is described by Matsomura [55]. Tabulated results show that the error squared obtained with adaptive control under various test conditions was between 11% and 46% of the error squared obtained with P1D control.
The main practical problem that [64] identified was that the plant needed to be persistently excited by superimposing a load test signal onto the load demand signal. This is needed for the identifier to adjust to different plant dynamics and parameters. It is not practical to have a power plant change its load continually just to update its controller parameters. It was proposed that the existing load demand signal may be used as the source of excitation if it is sufficiently rich in frequency to enable good plant identification. Matsumura [55] addressed the problem of persistent excitation by temporarily suspending the parameter estimation when the amplitudes or gradients of the input signals to the plant are small. 71
3.5.7 Adaptive control with prediction Valsalam [43] documented the application of a predictive controller to steam temperature regulation in thermal power plants. A model of the process was derived in the state-space form using the law of conservation of energy. The process model comprised a lumped parameter, discrete-time, second-order, state-space model combined with an empirical furnace heat transfer model calibrated at 100% and 60% load points. A Kalman filter [65] was used for optimal state estimation, filtering and prediction.
API controller was configured as an adaptive controller via on-line gain scheduling. The proportional and integral gains were computed from the discrete-time model of the superheater and the stability criterion in the Z-domain. The predicted steam temperature was used instead of the measured value for closed-loop control. In this way, the effect of the inherent process lag was nullified. Results obtained from simulation studies indicated an improvement in steam temperature control as a reduction of 10°C in the magnitude of temperature excursions.
3.5.8 Fuzzy logic control Fuzzy logic controllers are increasingly being used as nonlinear alternatives for PD, PI, and PID controllers [66]. The fuzzy logic algorithms are implemented as stand-alone, single loop controllers or as control modules in a DCS or PLC [67].
Although the rule-based of system fuzzy logic controllers have the ability to capture human expertise and deal with uncertainty on ill-defined systemS, its value to' he operation of well-characterized systems are less obvious [68]. This is backed up by practical experience obtained with fuzzy logic on nuclear reactor load control. The fuzzy controller had comparable accuracy to an analytical (model based) controller, had a slower response time and was more difficult to maintain under reactor refuelling conditions.
Fuzzy logic control was shown to perform better than conventional PI control on a simulated steam temperature control problem [69], but the results were below the standard attained with model-based feedforward control on the same problem. 72
3.5.9 Dynamic matrix control
The matrix controller has its origin in the petroleum refining industry, being applied to
distillation columns [70]. Its application is now wide-spread on many types of process
units. The controller addresses the problem of how to handle a process with several manipulated variables that affect several controlled process variables simultaneously.
Simple PID control usually provides inadequate performance due to the interaction
between loops. Often, the addition of specific feedforward compensatory loops can
decouple these interactions [42]. The matrix controller provides a unified approach to
multivariable control that replaces PID controllers and the decoupling feedforward loops.
All input-output interactions are considered.
The matrix controller functions by first predicting the future values of the controlled
process variables, with the assumption that the controls are frozen at their present values.
The objective is to determine error estimates that can be used to calculate the control actions needed to keep the process variables on setpoint. An error vector is derived for each process variable. The control action is determined by minimizing a cost function that contains the error vectors, the control actions, and weighting matrices. This makes it a class of optimal control.
Rovnak [70] documents the application of dynamic matrix control to a simulation of a supercritical thermal power plant. The objective was to control steam temperature, steam pressure and generated load by manipulating water flow, fuel flow, and goVemor valve position. Although documented results obtained, with matrix control shoW temperature deviations of -2.8°C to +4.4°C, no comparison is made with MD results. However, dynamic matrix control is shown to control the power plant satisfactorily.
3.5.10 Other techniques
In some cases it is possible to devise a control strategy based not on formal theory, but rather on operating experience and observations made during commissioning and testing.
In the case of Kendal Power Station it was noted that the reheater temperature deceases 73 sharply during load reductions. Since these units are operated in sliding pressure mode
(boiler pressure is changed in relation to boiler load), under-firing was excessively large to achieve a threefold objective:
Reduce boiler load Reduce the energy reserve stored in boiler pressure
Overcome thermal inertia in the boiler.
The pressure controller was then modified to decrease the extent of under-firing required by reducing the down-ramp rate of boiler pressure from the sliding pressure requirement to 0.1 MPa / min [71]. Due to operational difficulties, this limit on down-ramp rate was later increased to 0.25 MPa / min [44]. Temperature errors were decreased from -20°C to -10°C by this method.
Another technique was successfully implemented by Aitchison e.a. [39] after noticing that the largest steam temperature error occurred during start-up or shut-down of the third coal mill. This temperature deviation was caused by a sudden increase in the primary air flow. The error in temperature was reduced by momentarily opening the gas tempering damper and then slowly closing it, using a simple "kicker" circuit. The main steam temperature error was reduced from -9 and +9°C to -5 and +7°C in this way.
An interesting method of compensating for the large thermal lags of thick thermocouple pockets are described in [54]. Here the control system passes all thermocouple signals through phase lead compensators using time constants derived from plant tests and adjusted with boiler load. This compensation proved highly effective in improving the quality of control, particularly by its increase of speed of the inner desuperheater loops. 74
4. Neural networks and process control
4.1 Description of a neural network
4.1.1 Origin of neural networks A neural network (or more correctly, an artificial neural network) is a man-made system motivated by the neural structure observed in living organisms [72]. Mathematical models of biological neural networks created by neurophysiologists showed similar properties to those of the biological systems they described i.e. adaptability, learning, feature classification, and generalization of learning from past experiences to new experiences This led to the creation of electronic networks with the same structures observed in biological systems and hence the term artificial neural networks.
A neural network learns some input-output characteristic using a particular set of examples. Each example consists of an input pattern and a desired output pattern. It is postulated that neural networks mimic the way an animal learns and copes with an incomplete or confusing information set. Like an animal brain, a neural network can learn complex nonlinear relationships even when the input information is noisy and imprecise
4.1.2 Artificial neurons A neural network is composed of many simple and similar processing elements called artificial neurons (Figure 4.1).
Figure 4.1 Schematic representation of a typical artificial neuron. 75
The inputs of the neuron (x) may be externally measured variables or it may be the outputs of its own and / or other neurons. The output of the neuron (y) is a nonlinear function of
the weighed sum of its inputs. Neurons with no inputs and a constant unity output,
known as threshold or bias neurons, are implemented in a neural network where a
constant offset is required. Neurons of which neither the inputs nor the outputs are
externally connected are called hidden neurons.
Three kinds of transfer functions are commonly used in artificial neurons [75]: the hard-
limiter, the threshold, and the sigmoid (Figure 4.2).
Hard—limiter Threshold Sigmoid
Figure 4.2 Neuron transfer functions.
Assigning the sigmoid transfer function to a neuron is especially attractive due to its
continuity and boundedness and also due to the simplicity of calculating partial derivatives
through these neurons during the training of the network [76]. The function for a sigmoid
is given by:
1 fix) - (4.1) 1 +
4.1.3 Network Topology
The neurons in a neural network are arranged in layers and coupled together through 76 information-carrying connections. Two basic network topologies exist: feedforward and recurrent. The difference is that a neuron in a recurrent neural network may receive inputs from all neurons in the network including feedback from itself, where a neuron in a feedforward neural network may receive inputs only from the neurons in the preceding layer, or from the network inputs.
In a feedforward network the information is passed forward through the layers. The first layer is the input layer and it is provided with data obtained external to the network, i.e. plant measurements, calculations or data tables. Following the input layer, there are normally one or more hidden layers. Finally there is an output layer which present the desired data based on the inputs and the internal state of the neural network. Figure 4.3 illustrates a feedforward neural network with three inputs, two hidden layers with three neurons and a bias neuron in each layer, and an output layer with one neuron.
HIDDEN 1 HIDDEN 2
INPUTS OUTPUT
Yi
U,V,V denote layers • of weights.
BIAS NEURONS
Figure 4.3 Feedforward neural network.
The connections between the neurons in a neural network each have a certain internal gain called a weight. Changing the weight of a connection will alter the behaviour of a neuron, and therefore, it will also alter the behaviour of the whole network. The goal of training a neural network is to alter the weights in the network in such a way that the neural 77
network achieves the desired input / output relationship.
4.2 Selecting the size of a neural network •
A neural network must at least have an input and output layer. Hidden layers act as layers of absiraction and helps the neural network generalize results for inputs which it has not been explicitly trained on [77]. Increasing the number of hidden layers augments the processing power of the neural network but also significantly increases processing time and complicates training. It has been shown that a feedforward neural network with at least one hidden layer has the capability to approximate any desired nonlinear function to an arbitrary degree of accuracy [78]. Even though networks with only one hidden layer already have the desired approximation power, Draeger e.a. [79] report that two hidden layers give better convergence in the training process. A common method for determining the number of hidden layers is by experimentation [77]. However, due to the added processing burden, it is advisable to use more than one hidden layer only when it becomes necessary due to the inability of a a network with single hidden layer to train.
The number of input and output neurons are determined by the application of the neural network. Determining the number of neurons in the hidden layer is another experimental exercise [77]. Some rules of thumb have been said to give a starting point for estimating the number of hidden neurons: If one hidden layer is chosen, its number of neurons may be chosen as 34 the number of inputs to the network [80]. The number of hidden neurons should be equal to two times the square root of the number of input and output nodes summed [81]. The number of hidden neurons should total the number of training data sets divided by between two and ten times the sum of input and output nodes (ten for noisy data and two for clean data) [82]. Brainmaker documentation [83] also suggests estimating the number of hidden neurons by taking the average between number of inputs and outputs of the network.
These rules are silent on the complexity of the patterns in the training data. Since the neural 78 network should approximate some input-output relationship that could be nonlinear and multidimensional, the more complex the relationship, the more neurons are required. Too few neurons in the hidden layer may prevent the network from properly mapping the inputs to outputs.
On the other hand, too many nodes promotes memorization of specific input-output data points and inhibits generalization [77]. Memorization occurs when the patterns presented to the network are reproduced exactly without extracting the salient features. The network is then unable to process new patterns correctly because it has not discovered the proper relationships. Depending on the total number of nodes in a network, a sufficient number of training sets must be provided • in order to train the system adequately. Otherwise, the situation is the same as trying to fit a third- order polynomial to two data points, where an infinite number of sets of polynomial coefficients can satisfy the equation.
4.3 Training the network
Training a neural network is most commonly done through the error backpropagation method
[84]. Firstly, the inputs from a data set is presented to the neural network and the network outputs are calculated. The outputs are compared to the target outputs from the data set and the difference (error) is calculated. The backpropagation method assumes that all the neurons and connections are to some extent responsible for the error. It uses the chain rule of derivative calculus to allocate a portion of the error to every neuron in the network. The error is propagated backward from the output layer to the previous layer through the connections between the layers.
This process is repeated until the input layer is reached. The weights of the connections are then adjusted in the opposite direction of the partial derivative of the error. Many runs6f all the data sets are required during the training phase before weight convergence takes place.
The backpropagation algorithm for network weight adjustment is well known in neural network literature and will not be repeated here. (For example, [85] derives the backpropagation algorithm for minimizing the mean-square error of outputs by adjusting the weights for a network with a linear output layer.) However, some other aspects of backpropagation will be discussed later. 79
4.4 Process modelling with neural networks Many computer methods have been used for the simulation and control of power plants [86]. Statistical time series methods, involving the empirical fitting of parameters to autoregressive moving average models have been extensively used for plant simulation and dynamic matrix control, but require considerable competence in statistical methods [87]. Another method of modelling is the analytical approach where the entire model is based on physical properties and a set of mathematical equations. An example of the creation of an analytical power plant model is documented with results by Klefenz e.a. [88]. Other examples of power plant models are documented by March [52]. The generation of analytical models is labour intensive and these models have to be fine-tuned by adjustment of certain built-in factors [88]. Other methods including the nonlinear generation of empirical response surfaces [86], nonlinear regression [89], linear system identification [90], and fuzzy identification [91] have also been explored. Neural network technology offers an alternate method for the generation of process models. The advantages of using a neural network to represent a system are its ability to perform a nonlinear mapping between inputs and outputs and the necessaty of requiring minimal prior knowledge of the system.
It has already been said that a feedforward neural network with at least one hidden layer has the capability to approximate any desired nonlinear function to an arbitrary degree of accuracy. In fact, there is strong evidence to support that the learning mechanism of a neural network is simply a complex curve-fitting method that allows a network of simple processing elements to behave in a complex fashion [92]. The nonlinear mapping capabilities of neural networks allow the creation of accurate models of nonlinear processes. For example, one of the Most nonlinear industrial processes, being pH control in a neutralization tank, has successfully been modelled using a neural network [79]. Since neural networks produce an output pattern based on an input pattern and its prior training, they particurarly lend themselves to the modelling of complex systems.
Dynamic processes can also be modelled with neural networks. Dynamic process models require some kind of dynamic state feedback, either internal or external to the neural network. A neural network with feedback is termed a recurrent neural network and functions as a discrete-time 80 system model. Internal feedback is achieved by using the output of a neuron as an input to itself, or as an input to a neuron in a preceding layer [92]. External feedback is achieved by providing the neural network with inputs originating from previous outputs and previous plant states [93].
The main function of the feedback in the neural network is to encode a time-based memory into the network.
The external feedback method was used to model a 200MW power plant unit at Ballylumford power station in Northern Ireland [94]. The neural network power plant model had 16 inpUts,
24 hidden neurons and 4 output neurons producing the 4 modelled outputs. The 16 inputs consisted of 4 manipulated variables and their values delayed by one time step, as well as the 4 previous outputs of the model and their values delayed by one time step. The network was trained on noisy data from a validated computer simulation. The results obtained with the neural network model were comparable to those obtained with a linear multivariable autoregression model at two predefined operating points. The neural network model was shown to produce significantly improved results of the plant outputs across the complete operating range. A similar exercise was done by Reinschmidt [86], who also achieved a very accurate power boiler model.
Mother example of a dynamic model of power plant systems is the modelling of the evaporator and steam drum of a 235 MW Clifford B. Jones unit by means of a neural network [95]. The model comprised three task-specific neural networks that were configured with external feedback.
Training data was obtained from a plant simulator developed previously. Results obtained with the neural network model were compared to output data from the simulator and showed good drum pressure and drum level modelling.
The modelling capabilities of neural networks have also been proposed for inferential sensors to obtain estimates of various process variables for which no easy method of on-line measurement exists [96]. Also called soft sensors, these neural network based virtual instruments have been applied with great success to industrial processes [97], paper making machines [98], and power boilers [99], while user configuration makes these systems capable of inferring many unmeasured variables on-line [100]. 81
Neural networks are not limited to simulation in the sense of predicting the response to a specified action, but can also be used to generate the action necessary to produce a given response, i.e. to control a process.
4.5 Process control with neural networks Much research is being done in the field of process control via neural networks, and the general suitability of neural networks for control purposes has already been demonstrated [84]. Neural network controllers are becoming commercially available, for example NO x emission control and boiler efficiency improvement [99] & [101]. There are basically five generic designs for using neural networks to directly control processes of some kind [76] & [102]. Some hybrid control designs have also been proposed.
4.5.1 Supervised control In supervised control, a neural net learns the mapping from sensor inputs to desired actions, by adapting to a training set of examples of what it should have done. Thus one can "clone" a human expert or some other control system with a neural network controller. The disadvantage of this technique is that the neural network control will only match, but never surpass, the,control quality of the human or initial control system [103].
The supervised control performance of a recurrent neural network on controlling the temperature of a batch reactor in real-time was evaluated by Dirion e.a. [104]. In two experiments, the neural network controller was trained on control actions produced by an adaptive controller and on human control actions. In both cases the neural network controller was able to learn and mimic the original control actions. It also maintained satisfactory control under situations that were not part of the training set.
4.5.2 Neural adaptive control In neural adaptive control, linear mappings used in standard designs such as Model Reference Adaptive Control are replaced by neural nets, resulting in greater robustness and greater ability to handle nonlinearity [105]. As in all adaptive control techniques, the neural adaptive scheme comprises identification and control performed by an on-line 82 adaptive structure. The design is based on two neural networks. The first learns the unknown dynamics of the plant. The second uses this knowledge to adjust its connection weights and to generate the control signal on-line [106]. The ability of neural networks with a dynamic learning algorithm to model arbitrary dynamic nonlinear systems makes the control scheme less sensitive to variations in system parameters [107].
This technique was applied by Khalid e.a. [108] to control the temperature of a laboratory water bath. It was shown that the performance of the neural network controller was superior to that of a PI controller under the influence of load disturbances and varying plant dynamics. Furthermore, [109] demonstrated the inherent capability of a neural network-based adaptive controller to handle nonlinearities, learn, and perform control effectively for a real-world system, based on minimal system information.
4.5.3 Adaptive critic Adaptive critic methods show promise in reproducing the self-learning capabilities of the animal brain by exploring the effect of new and different control actions. The underlying concept is to add a random bias to the output of a neural network controller and if the control action is better than expected, the controller is trained to reproduce the "new" action given the same inputs [110].
4.5.4 Direct inverse control In direct inverse control, a neural net learns the inverse dynamics of a system! By applying the desired range of inputs to the plant, its corresponding outputs can be recorded and a set of training patterns can be obtained. Once trained, the neural network uses the desired system state as inputs and the network output becomes the control input to the process. Sbarbo-Hover e.a. [111] demonstrate how this technique could be applied to a steel rolling mill. Results obtained on a plant simulation showed a marked improvement over PI control.
4.5.5 Back propagation through time The back propagation through time scheme adapts a controller by solving a calculus of 83
variations problem. This scheme has been applied in situations where direct inverse dynamics will not work because of system singularities [76]. As with the calculus of variations, this method requires a model of the system to be controlled. By propagating the output error backwards through the model, it can be determined what the error on the input of the process was, and the controller can be adjusted or trained accordingly. The backpropagation technique calculates the derivative of an error on the output with respect to the inputs.
Derivative of errors with respect to inputs If a process model has a defined inverse, control actions for reducing the errors could be calculated without much effort. However, in the case of the boiler under consideration in this thesis, no defined inverse of the process exists. It will be shown later that the boiler model maps seven inputs (5 mill fuel flows, air flow index, and burner tilt angle) onto three outputs, resulting in an infinite number of possible input (furnace element) configurations that may produce the same output (heat transfer) pattern. Also, many output heat transfer rates cannot be achieved, regardless of the input conditions. The • control signals for minimising the errors must therefore be calculated in some other way.
Perturbation of process inputs Facing a similar problem with the optimization of synthetic fuel reactor production [112], a neural network system was designed to allow controlled variables, u, to be perturbed as a means of establishing the best values of the input variables. Each input to the neural network reactor model was varied in small amounts and then adjusted according to the direction of change observed on the network output. This method actually determined the partial derivative of the network output with respect to the inputs. Adjustment to the inputs were based on the sign of the partial derivative.
Backpropagation of error Another, more elegant method, was used by Werbos [113] for optimizing long term gas industry profits. The technique, utilizing backpropagation, is essentially just a variant of the steepest gradient method for minimizing or maximizing functions. When a neural 84 network is used to represent a system, the backpropagation algorithm can be used to propagate the error derivatives backward through the model (Figure 4.4), eliminating the explicit calculation of the Jacobians of the model [114]. Once the errors have been backpropagated through the neural network model to appear on the model inputs, these are used as equivalent errors on the controller outputs. The weights of a neural network controller can then be adjusted to minimize the equivalent errors.
The backpropagation of error technique in a sense translates the error in the plant output to the error in the controller output [115]. The real plant cannot be used here because the error cannot be propagated through it. The relative simplicity of the backpropagation algorithm is good motivation for using neural networks for plant modelling in place of analytical methods. For the control of dynamic systems, a run of control actions and plant outputs are recorded over a predefined number of time steps. The backpropagation technique is then applied recursively to every time step of the recorded run, starting with the last run [116]. The error is propagated backward through time, hence the name of the technique, backpropagation through time.
• Previous Controller Plant model output Error -411( Backpropagation Backpropagation Setpoint and adjustment of error of weights
Plant Control signal Output
Setpoint
Figure 4.4 Backpropagation signal flow.
Mechanics of Backpropagation Werbos [116] states that "Backpropagation" refers to how the derivatives of a neural network map is calculated and has nothing to do with errors. However, to have purely 85
a derivative, it will be sufficient to backpropagate unity, but since the backpropagation technique is used for adjusting parameters, it is useful to backpropagate the size and sign of the error at the same time. This saves the effort of calculating the derivative first, and then calculating the adjustments, based on the derivative and the size of the error. Backpropagating the error does both calculations in one pass.
To demonstrate the difference between the normal feedforward operation of a neuron and backpropagation, consider the neuron in Figure 4.5. In the feedforward mode, the algorithm takes the sum of the inputs x, and passes it through the sigmoid function to obtain the output y. In backpropagation mode, the error e is multiplied by the derivative of the sigmoid, y (I -y), to obtain the equivalent error 4 which is then multiplied by the appropriate connection weight. Equivalent errors backpropagated from different parts of the network and arriving at the same node are simply added together.
-401111111111111111111111111111111111111111111 w, Backpropagation O
y - e
X=E x, O - x, Feedforward 11111 111■11■MOIO.-
Figure 4.5 Feedforward and backpropagation modes.
By using the technique of backpropagation, it is possible to calculate the partial derivative of any network output with respect to any network input. This information can then be used to calculate adjustments needed on the inputs that will change the outputs in such a way that some cost function .1 is minimized. In order to do this, the backpropagation 86 algorithm calculates the partial derivative of J with respect to every input. Once the partial derivative of the output to an input is known, adjustments can be made to the controller.
This technique lends itself very well to neural network control, because with backpropagation through time, a neural network controller can be trained without much prior knowledge of the system to be controlled. No training data from other controllers to be replaced by the neural network is necessary because the training is done on data captured on-line. By virtue of providing the partial derivatives of process outputs with respect to process inputs, the .backpropagation algorithm could lend itself well to solving optimization problems too. This aspect will be explored in more depth later, where it will form the basis of a new steam temperature controller.
4.5.6 Hybrid Neural Designs Other techniques have been proposed that use the modelling ability of neural networks. These schemes are based on other advanced control methods but use a neural network in some part of the design. Many of these hybrid neural designs exist, but some examples are: generalised minimum variance control [117]; neural predictive control [77] & [118]; optimal control [86] & [111], and neuro-fuzzy 'control [119]. 87
5. Plant modelling
The new steam temperature control strategy developed in this research project, uses a neural network model of the heat transfer between furnace and boiler elements as its core. Although the controller design and its structure should ideally be discussed before the details of any of its components, the author chose to discuss the model before the controller design. The reason is that observations made during the modelling phase determined many aspects of the controller design and to save discussing these issues twice, modelling will be dealt with first.
5.1 Desired model characteristics
The objective of the heat transfer model is to map the furnace conditions to heat pickup in the boiler. Judgement on the inputs to use, the nature of the test data, and the structure of the model was made on past experience and heat transfer theory. The key considerations are discussed below.
5.1.1 Individual mill firing rate
Experience has shown that the bottom mills are more suited to producing pressure and the
top mills are more suited to producing temperature. When the bottom mill (E-Mill) is
placed into service a pressure excursion is likely to follow, while placing the top mill (A-
MID in service, the effect is predominantly seen on superheater temperature. Due to the
large burner spacing, the bottom mills discharge most of their heat onto the water walls
of the boiler, thereby producing evaporation, steam flow, and pressure, while heat
discharged from the top mills primarily increases the furnace exit tempirature, thereby raising the steam temperature.
Due to the effect that different mills has on heat transfer, it will not be sufficient to use
only the total fuel flow as an input to the proposed heat transfer model. For this reason,
fuel flow rates from each individual mill were used as inputs.
5.1.2 Burner tilt angle
The effect of burner tilt angle on heat distribution has already been discussed. Tilting the 88
burners downward increases heat transfer to the evaporator while decreasing heat transfer
to the radiant superheater (Kendal has virtually no radiant reheater surface). Tilting the burners upward has the opposite effect. Having a large effect on heat transfer necessitates
the incorporation of burner tilt angle as an input into the heat transfer model.
5.1.3 Furnace air flow
The influence of excess air flow on reheater temperature has already been established
during the commissioning of the Kendal units. The effect is so intense that it was tested as a primary control element for reheater temperature at one stage, but it is utilized as a
secondary control element presently. The relation between air flow and reheat temperature is due to an increase in furnace air flow leading to an increase in convective heat transfer and a reduction in radiant heat transfer. Adjusting the furnace air flow alters the distribution of heat between the evaporator, superheater and reheater due to the evaporator having mainly radiant surface, the superheater having both radiant and convective surface, and the reheater having mainly convective surface.
To replicate this effect, the heat transfer model was provided with an index of furnace air flow as an input. Although it is intuitive to use the total air flow measurement for this purpose, it will be shown later that the oxygen concentration in the flue gas was favoured above the air flow signal.
5.1.4 Windbox damper position
The secondary air flow into the furnace is controlled via dampers situated -in the windbox of each burner. These dampers can control the distribution of secondary air flow into the furnace, which should have an effect on heat distribution. Since the damper control philosophy has been established practically, based on combustion and flame stability observations, the dampers operate in a fixed manner. Due to this, the distribution of secondary air is always repeatable and may be considered as part of the furnace characteristics. Windbox damper positions were therefore not used as inputs into the heat distribution model. 89
5.1.5 Heat transfer rate
The outputs of the heat transfer model were heat transfer rates to the economizer,
evaporator, superheater and reheater. The heat transfer to these boiler components was determined by calculating the heat gain across these components.
5.2 Acquiring test data
As mentioned before, neural networks need to be trained. This requires that training data consisting of input patterns and the required output patterns be made available to a network in training mode. To obtain data for training the neural network model, a series of steady state tests were performed on Kendal Unit 3. This section describes the test objectives, development of the test programme, and running the actual tests.
5.2.1 Objectives of steady state tests
The objective of the steady state tests was to obtain data for the creation of a nonlinear
mapping between conditions on the fire side of the boiler and the heat pickup of different
boiler components. The tests were intended to be steady state tests and all data was
recorded under stable boiler conditions.
5.2.2 Covering the operating envelope
This test data had to be sufficiently rich in heat transfer characteristics so that the neural
network trained on it will be able to predict heat transfer rates outside the normal
operating regime (for example, with mills biassed). All the furnace elementlenvisaged as
model inputs were manipulated at different loads. The ranges of the manipulated elements . are listed below:
Boiler load (286 MW to 700 MW)
Mills in service (2 mills to 5 mills, all possible combinations)
Individual mill loading (40 % to 110 % mill fuel flow)
Burner tilt position (-30° to 30°)
Furnace draft (2.5% to 6% 0 2 in flue gas)
Before setting up a test programme, All of the mill combinations possible with 2 - 5 mills 90
were listed. The mill combinations were individually assessed on the basis of flame stability and fireball height in the furnace. Those combinations deemed to be unsafe, or bad operating practice were eliminated. Sixteen mill combinations remained. It was decided to do tests with each of the remaining mill combinations and test numbers were assigned (Table 5.1).
Mill combination Possible and safe? Test no. ABCDE. Yes 1 ABCD Yes 2 ABCE Yes 3 ABDE Yes 4 ACDE Yes BCDE Yes 6 ABC Superheater overheating ABD Dangerous if D-mill. trips ABE Double gap between B&E ACD Yes 7 ACE 3 unsupported mills ADE Double gap between A&D • BCD Yes 8 BCE Yes 9 BDE Yes 10 CDE Yes 11 AB Superheater overheating AC Superheater overheating AD Double gap between A&D AE Triple gap between A&E BC Yes 12 BD Yes 13 BE Double gap between B&E CD Yes 14 CE Yes 15 DE Yes 16 Table 5.1 Elimination of mill combinations.
It was decided to run the tests over a period of sixteen days, each day with a different mill combination. A test period of eight hours was planned for each day. During the eight hours, eight sub-tests could be run, each a duration of an hour, and each with a different 91
set of furnace conditions (i.e. unit load, burner tilt angle, % 0 2, and mill biassing). A total of 128 tests was planned for in this way.
During the hour assigned to each sub-test, the boiler was set up in the first fifteen minutes, left to settle out for thirty minutes, and the last fifteen minutes were used to record the steady state data. For each sub-test, one of the mill fuel controllers, the burner tilt controller, and the 0 2 setpoint generator were placed in manual mode and adjusted to a specific predefined value.
As an infinite combination of possible furnace / boiler conditions exist, a series of tests such as this one can only explore a very limited number of conditions. To ensure that the tests cover an even spread across all possible furnace / boiler conditions, all the predefined setpoints of the control elements (except of course mill combination) were chosen randomly. Firstly, upper and lower limits were placed on all control elements. The limits on loading of the biassed mill required careful consideration, because not only is the mill load restricted, but so is the compensating movement required from the other mills in service at the time. The unit load setpoint was first chosen randomly, then the mill to be biassed, then this mill's load setpoint. Thereafter, random settings were generated for 0 2 concentration in flue gas and burner tilt angle. The tests for each day were sorted in descending order of unit load level, to reduce the magnitude of load changes between tests and therefore reduce the boiler setup and settling times. A list of the various test conditions is attached in Appendix A.
5.2.3 Process data recorded Process measurements were recorded to capture the furnace conditions during each test and to enable the calculation of heat absorbed in the economizer, evaporator, superheater and reheater. Ninety-five process variables were recorded (see Appendix B). The number of recorded points is in line with similar documented boiler modelling, i.e. Zhu e.a. [45] recorded 100 data points for boiler plant modelling. Refer to Figure 5.1 and Figure 5.2 for schematic diagrams of the feed water system and boiler components indicating the location of the most important measured process variables (the symbol P denotes a point 92
of pressure measurement, T denotes temperature, and F denotes flow). Test data was recorded on the process computer at 5 second intervals over the last 15 minutes of a test. All process measurements were sampled via analog to digital converters with an effective resolution of 14 bits.
Deaerator Boiler Feed water Cold reheat storage feed regulating extraction tank pumps valves P,T Drum Economizer [P,T ei-pr T T To super heater PA Feed water heaters
Distillate Superheater Reheater y spray water Boiler spray water water circulating pumps
Evaporator From LP heaters
Figure 5.1 Measurements on feed water system, economizer and evaporator.
5.2.4 Operational requirements The following list of operating requirements was drawn up from a plant health, test integrity and plant safety point-of-view.
The permissible steam and metal temperatures were adhered to at all times. A minimum steam flow of 40% (230 kg/s) was adhered to at all times. The tests could be suspended at the request of the national load coordinator The entire boiler and furnace were soot-blown prior to testing. 93 e) The unit could not supply the auxiliary steam range. 0 Demineralised water make up may not exceed a daily average of 5 m ;/h. g) During all tests, automatic dispatch mode and frequency control was switched off.
Superheater Reheater spray water spray water
Superheater T Reheater
T,P P From To IP steam Primary Final turbine drum T, P T,P,F P
Steam 11 V extraction to Gland Reheater Superheater feed water steam spray water spray water heaters leakage
Figure 5.2 Measurements on superheater and reheater.
5.2.5 Performing the heat distribution tests
As planned, one hour was allocated to each test. The boiler was set up in the first 15 to
25 minutes after which time was allowed for firing rate and steam temperatures to settle
out.
Initially, some difficulty was experienced with extracting the data from the process
computer and a large backlog of data accumulated over the first three days. It was decided
to postpone the first weekend's tests to the next week to allow time for the computer
personnel to clear the backlog.
At times the tests were suspended by the national dispatch control centre (National
Control) who requested full load from the unit due to power shortages on the system. As 94 the test loads became lower (3-mill tests), it became progressively more difficult to obtain access from National Control for testing. Eventually, two days were lost due to National Control not granting access to do the tests, due to high system demands. It was decided to continue with the tests during the night - when the demand for load was less. Almost no access problems were experienced during night testing.
Test 1.7 was done at full load due to power system requirements. The final test for the day, Test 1.8 was cancelled on National Control's request. The last test on the third day, Test 3.8 was requested to be done at 430 MW and not at 402 MW as planned. A ninth (unplanned) test was done on the same day, also at 430 MW. Test 7.1 was stopped due to the very high demands it placed on the mills and consequent fuel and pressure cycling that occurred on the unit.
While doing the 11th set of tests, the mills on the test unit started choking and blocking due to coal that was wet as a result of an unexpected high rainfall. The tests were suspended after the third test and was resumed only after seven days due to these unfavourable wet coal conditions. When testing was resumed, the first three tests of day 11 were repeated and the remaining tests were run without problems. The series of tests were completed on 15 March 1996. In total, 130 tests were done.
5.2.6 Data processing and verification As said before, the process parameters for each test were recorded at 5 second intervals over a period of 15 minutes after the boiler conditions had stabilised.' The data was recorded on the process computer, from which it was transferred to a file server on the station Local Area Network (LAN) via a serial communications link. From the station LAN, the data was downloaded onto a Personal Computer (PC), and imported into Quattro Pro spreadsheets [120], one spreadsheet for each test. A total of 2.2 million data values were downloaded.
Verifying steady state condition One of the requirements of the steady state tests was that all the test data had to be 95 captured under steady boiler conditions. This was verified for every test by making a plot of key indicators of state over the 15 minute period of data recording. The variables plot in this way were: fuel flow air flow feed water and steam flow superheater and reheater spray water flow boiler pressure final steam temperatures on superheater and reheater.
Although minor fluctuations were present in the data, the boiler had settled out prior to the start of recording in all tests and all the data was deemed representative of the boiler under steady state conditions.
During the data verification phase, it was noticed that only nine minutes of data were downloaded for Test 1.4 and only five minutes of data for Test 2.7. However, due to the steady state conditions that prevailed during data capture, the test data was in both cases accepted as valid data. It was also noticed that some of the data points in Test 4.3 were missing. Consequently, the entire data set from this test was rejected, bringing the number of tests with valid data down to 129.
Average values Once the data for each test was deemed representative of steady state'conditions, the average of each data point was calculated. All the averages were compiled into a single spreadsheet. The heat transfer was calculated using these average values of temperatures, pressures and flow rates.
Calculating heat transfer Heat transfer to all the boiler elements was calculated by calculating the difference in enthalpy across an element and multiplying this with the flow rate through the element. These calculations required the steam or water enthalpy at 26 positions between the 96 deaerator storage tank outlet and the IP turbine inlet for each of the 129 valid tests.
To calculate all the enthalpy (3354 in total), a special programme was written in C-H- [121] to calculate the steam and water enthalpy. The calculations were based on the IFC formulations of the thermodynamic properties of water for industrial use [122]. Spread sheet columns containing pressure and temperature measurements were exported to this programme, which then calculated the enthalpy of each pressure-temperature pair. Enthalpy of boiling water and saturated steam were calculated from either temperature or pressure. The set of enthalpy values was imported back into the spreadsheet where it was used for calculating heat transfer rates to the various boiler components.
Independent calculations of heat transfer rate were done for all the following components: Economizer Evaporator Left-hand primary superheater Right-hand primary superheater Left-hand secondary superheater Right-hand secondary superheater Left-hand final superheater Right-hand final superheater Left-hand reheater Right-hand reheater
Test data integrity After calculating the heat transfer to all the individual components, the integrity of the data was analysed by graphically comparing related variables to each other. For all the tests, the following variables were compared graphically for linear relationships: measured spray flow against calculated spray flow generator load against fuel flow measured air-fuel ratio against calculated air-fuel ratio total heat transfer against fuel flow. 97
The first three graphs are discussed in the next section. Figure 5.3 shows the correlation between total heat transfer and fuel flow of the 129 tests. The plot of total heat transfer rate against fuel flow would have indicated any tests containing corrupt data sets.
1800
1600
2 1400
1 1200
et 1000
800 it- 600 40 50 60 70 80 90 100 110 Fuel flow [%]
Figure 5.3 Correlation between fuel flow and total heat gain was obtained for all tests.
Initial test results Heat transfer to the main boiler elements were Charted against steam flow to indicate the heat shifting potential of the furnace (Figure 5.4). Through manipulation of the furnace elements, the following heat shifts (away from the average) were obtained during the heat distribution tests:
Economizer & Evaporator: +10%, -10% Superheater: +20%, -20% Reheater: +30%, -20%
At this stage, it was already obvious that it was indeed possible to manipulate the distribution of heat between the different boiler elements by adjusting the furnace elements. It was expected that even larger heat shifts could be achieved if all furnace elements were adjusted in tandem to obtain a specific effect. 98
800
600 ■■tliii■■■■■■ PIES2111:1■1■■
200 MaitiliNININGPATal
0 200 250 300 350 400 450 500 550 600 Steam Flow [kg/s] Lo Eco + Evap x Superheater o Reheater Figure 5.4 Heat shifts achieved during heat distribution tests. Straight lines represent a least squares linear fit.
5.3 Calculations and assumptions To keep costs of this project to a minimum, no additional instrumentation could be installed on the test unit. Even with this limitation, it was possible to obtain all the necessary variables needed for calculating the heat transfer. Where possible, measurements of the variables were obtained directly from existing sensors and transmitters on the plant. Where the measurements were not available, variables were obtained indirectly from calculations based on plant measurements. This section describes these calculations and any assumptions that were made are described and motivated here. The limitations on certain measurements are explained and any discrepancies in the results of the heat distribution tests are discussed.
5.3.1 Burner Tilt Positioning The Kendal burners are tilted via a pneumatic power cylinder controlled by pneumatic positioners with mechanical position feedback. These units have been found to be susceptible to calibration shifts which affect the actual tilt angle. The true burner tilt angle is not electronically fed back to the boiler control system. Also, if a burner becomes 99
mechanically seized, the fault is not detected by the control equipment.
During the period of testing, burner A-2 was stuck at -22.5° for any setpoints greater than
-22.5°. Some burners had large deviations from setpoint and virtually all burners were not
achieving the full ±30° travel. Additional technician assistance was requested to check and
recalibrate burner tilt positioners with large offsets. Examples of three typical sets of
physical burner tilt angles for high, horizontal, and low tilt setpoints are shown below.
CNR1 CNR2 CNR3 CNR4 A -18 -26 -26 -26 B -24 -25 -26 -10 C -24 -26 -26 -24 D -15 -15 -7.5 -7.5 E -22.5 -20 -26 -17 Table 5.2 Tilt performance: setpoint = -28°, average angle = -21°
CNR1 CNR2 CNR3 CNR4 A 6 -22.5 5 3 B 3 0 2 5 C 2 4 0 3 D ,3 2 3 0 E 3 2 -6 5 Table 5.3 Tilt performance: setpoint = 0°, average angle = 1°
CNR 1 CNR 2 CNR 3 CNR 4..0 Q 27 -22.5 7.5 27 co 20 28 15 20 o 24 26 22.5 28 0
W 23 28 29 10
25 26 17 .28 Table 5.4 Tilt performance: setpoint = 30°, average angle = 20°
To obtain representative data, the actual position of every burner in service was noted
during a plant inspection done as part of each test. The burner tilt angle used for modelling was taken as the average of all the individual burner angles. 100
5.3.2 Deaerator storage tank enthalpy Some of the pressure and temperature measurements of the water inside the deaerator storage tank converted to slightly superheated steam enthalpies. This could be expected, since the vessel contains boiling water and saturated steam and small deviations on the temperature and pressure measurements could very well indicate compressed water or superheated steam. This problem could be overcome by using either one of the measurements and assuming boiling conditions. It was decided to use the temperature measurement to calculate the enthalpy of the water in the deaerator storage tank.
5.3.3 Reheater spray water enthalpy Due to the lower pressure of the reheater compared to the superheater, the reheater spray water is extracted from the second stage of the main boiler feed pumps. The pressure and temperature of the extraction are less than that of the pump discharge, and so will be the enthalpy. Extraction temperature and pressure measurements were not available on the plant, and so the actual enthalpy of the reheater spray water could not be determined from measurements.
Plant measurements were available for calculating the inlet and discharge enthalpy of the boiler feed pumps. Design heat flow diagrams prOduced by the turbine manufacturer show a linear relationship between total enthalpy rise over the boiler feed pumps and enthalpy rise from inlet to reheat spray water extraction [123]. These design sheets show that 36.9% of the enthalpy rise takes place before the spray water extraction point and the rest thereafter, regardless of boiler load. It was assumed that these design calculations hold true for the actual plant.
5.3.4 Enthalpy loss in spray water lines As the hot spray water is transported along the piping between the boiler feed pumps and the spray water injection points, some heat loss will occur. No measurements were available to obtain the actual spray water enthalpy before injection. The spray water lines are clad with thermal insulation to keep the heat loss to a minimum, and the flow rate through these lines are quite high, so the decrease in enthalpy can be expected to be small. 101
However, it is necessary to estimate the impact of heat loss on the spray water enthalpy.
The surface temperature of the superheater spray water pipe was measured through a small hole in the thermal cladding at the boiler feed pump and at the desuperheater by means of an infrared thermometer. The decrease in pressure along the line was assumed to be 1.1 MPa (although pressure has very little influence on the enthalpy of water). The calculated enthalpy of the superheater spray water at the two positions is shown in Table 5.5.
Position Pressure Temperature Enthalpy [MPa] [t] [kJ/kg]
Boiler feed pump discharge 20.1 177 760.2
Desuperheater inlet 19.0 170 729.4
Table 5.5 Superheater spray water enthalpy.
Under this assumption, a decrease in spray water enthalpy of 30 kJ/kg occurred along the spray water supply line. Since the spray water is heated and evaporated inside the desuperheater, it is possible to express the heat loss in the piping as a percentage of the heat of absorbed by the spray water. The enthalpy loss in the pipe equates to 1.5 % of the heat absorbed in the desuperheater.
Thus, by neglecting the effect of heat loss in the spray water piping, an error of about 1.5 % will be induced in spray water flow calculations based on a heat bakince across the desuperheater. This error on spray water flow rate is negligibly small in comparison to the main steam flow rate. Because the heat loss in the superheater spray water supply piping cannot be accurately determined and due to the very small effect on the process as a whole, it was ignored in desuperheater heat balance calculations. On the same grounds, heat loss in the reheater spray water supply piping was ignored.
5.3.5 Main steam flow balance The main steam flow rate signal recorded during the tests, is derived from the pressure before the first stage blading on the HP turbine via a choked gas-flow calculation. 102
Therefore, it is not possible to obtain the main steam flow rate as individual left-hand and right-hand flow rates from the recorded data. Knowing the steam flow on each of the two individual flow paths is a prerequisite for making independent spray water flow rate calculations for each desuperheater and it is also essential for heat transfer comparisons between the two sides of the superheater.
It is natural to assume equal steam flow rates on the two sides of the superheater, but if this assumption is false, large errors could be made in terms of heat transfer calculations and spray water calculations. However, at Kendal, a backup steam flow measurement is installed. This second steam flow measurement is based on the differential pressure across the final stage of the superheater. As separate measurements exist for the left-hand and right-hand superheater, these were used after the tests to compare the steam flow rate through the two sides of the superheater.
No noticeable difference in steam flow rate existed between the left-hand and right-hand superheater. It was therefore assumed that the steam flow rate at each superheater outlet is equal to the main steam flow rate (as calculated from the steam pressure before the turbine blading) divided by two.
5.3.6 Reheater steam flow Reheater steam flow rate is not measured at all. This flow differs from the main steam flow due to steam leakage past the HP turbine gland seals, and also due to the extraction of steam from the HP turbine exhaust. The steam extraction is used to heat the feed water as part of the regenerative Rankine cycle.
There was no plant instrumentation to measure the steam leakage rate or other measurements from which this value can be calculated. Therefore, the steam leakage rate was estimated from values indicated on the turbine heat flow diagrams [123]. The following simple linear relation between main steam flow and the design steam leakage was established: M = 0.005194 mans (5.1) 103 where: steam leakage rate ma, = main steam flow rate
To calculate the extraction steam flow rate, a heat balance calculation was done across the feed water heater (Figure 5.5).
Extracted steam m r• , hUe
Feed water inlet Feed water outlet m • h .11
Distillate
h d
Figure 5.5 Feed water heater.
The feed flow enters the heater, and is heated by the extracted steam. The heat balance across the heater is described by Equation 5.2.
ma (ha - 12,1) = mf (hro - hfi) (5.2) where: ma = extraction steam mass flow rate mf feed water mass flow rate ha = extraction steam enthalpy hd = distillate (condensed extracted steam) enthalpy hfi = feed water enthalpy at heater inlet feed water enthalpy at heater outlet 104
Equation (5.2) may be rewritten to obtain the extraction steam flow rate:
mf (ho - hfi) m (5.3)
(her - hd)
Due to the physical location of the available temperature and pressure sensors on the plant,
not all temperature and pressure measurements were available for solving Equation (5.3).
Therefore, the following assumptions were made:
The extraction steam pressure and temperature measurements were taken at the
turbine exhaust and not at the heater inlet. It wass assumed that no loss of
enthalpy occurred in the pipe between the turbine and the heater. This assumption
is supported in design sheets produced by the turbine manufacturer (Table 5.6).
Load Position Pressure Temperature Enthalpy
[MPa] [°C] [kJ/kg]
40 % Turbine exhaust 1.6148 331.3 3105.5 .
Heater inlet 1.5629 330.8 3105.5
100 % Turbine exhaust • 4.0997 331.2 3044.5
Heater inlet 3.8946 329.1 3044.5
Table 5.6 Turbine outlet and feed water heater inlet conditions. [123]
The distillate outlet pressure was not measured. Because of its small effect on the
enthalpy of water, it was assumed that the pressure difference between the steam
inlet and distillate outlet is negligible so that the extraction steam pressure may be used for enthalpy calculations. Table 5/ indicates virtually no difference in
distillate enthalpy at the minimum and maximum pressures possible for the design
outlet temperature. Design temperatures were obtained from [123]. No distillate
outlet pressures are stated in the mentioned source. 105
Load Distillate state Pressure Temperature Enthalpy [MPa] [t] [kJ/kg] 40 % Compressed liquid 1.5629 164.1 693.83 Saturated liquid 0.6853 164.1 693.33 100 % Compressed liquid 3.8946 205.2 876.71 Saturated liquid 1.7314 205.2 875.89 Table 5.7 Distillate conditions.
c) The feed water inlet and outlet pressures were not measured, but the discharge pressure of the feed water regulating valves upstream of the heaters was measured. As with the distillate, it was assumed that the pressure difference across the heater has a negligible effect on the enthalpy of the feed water. The measured feed water pressure at the feed water regulating valve outlets could therefore be used to calculate the feed water enthalpy at the heater discharge. Design pressure differentials were obtained from the turbine design heat flow diagrams [123] and the effect on enthalpy is negligable, as demonstrated in Table 5.8.
Load Feed water conditiOn Pressure - Temperature Enthalpy at heater discharge [MPa] [°C] [kJ/kg] 40 % Actual [123] 19.836 247.0 1072.6
Pressure = inlet 20.106 247.0 . :1072.7 100 % Actual [123] 8.620 204.0 873.0 Pressure = inlet 8.680 204.0 837.2
Table 5.8 Feed water discharge conditions.
Having made the three assumptions, the steam extraction rate could be calculated. Then the steam extraction rate and the gland steam leakage were known and could be subtracted from the measured main steam flow rate to obtain the flow rate of the cold reheat steam.
Similar to the superheater, the reheater is also divided into a left-hand and right-hand side. 106
No measurements on the plant existed to determine the flow distribution between the two parts of the reheater. Based on the mechanical equivalence of the two sides of the reheater, the assumption was made that each side carried an equal part of the total reheat steam flow rate. The flow rate entering any one side of the reheater was set equal to one half of the total cold reheat flow.
5.3.7 Steam pressure measurement Steam pressure was measured at the superheater inlet (steam drum) and at the superheater outlet. It is necessary that the steam pressure is known at each of the desuperheaters to be able to calculate the steam enthalpy for heat balance calculations. But steam pressure was not measured at the desuperheaters. Only a measurement of the pressure differential across the final stage of the superheater existed additional to the drum pressure and the final steam pressure.
0.25 0_ X — 0.2 X
a) X X
X 0.15 x X
To 0.1 Lc:
20 0.05 ca a_ 0 0 1 J t I I I 1 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 DP across entire superheater [MPa]
Figure 5.6 Relation in pressure differential (DP) across superheater stages.
The pressure differential across the final superheater stage was measured and compared to the total pressure differential across the entire superheater. A zero-zero origin was 107
assumed, related to the condition of no steam flow. A least squares gradient fit was performed and a linear relation was found between the two pressure differentials (see Figure 5.6).
With minor variations, 15% of the total pressure differential across the superheater occurs in the final stage. The other 85% of the total pressure differential must then occur in the primary superheater and secondary superheater. (The primary superheater refers to the combination of all superheater stages before the primary desuperheater. The secondary superheater refers to all superheater stages positioned between the primary desuperheater and secondary desuperheater.)
Based on the linear relation between pressure differential in the final superheater stage and total superheater, it was assumed that the pressure differential across all superheater stages had a linear relation with total superheater pressure differential. The remaining 85% of pressure differential occurring in the primary and secondary superheater stages was assumed to be divided equally between these two stages, or 42.5% of the total pressure differential per stage. If this assumption was not true, the enthalpy calculations at the primary desuperheater would be less accurate. To test the extremes of the error possible, one could argue that the pressure differential across the primary superheater can only lie within 0% and 85% of total superheater pressure differential. If the pressure differential across the primary superheater was one of these extremes, and not the assumed 42.5%, the errors will be the largest (Table 5.9).
Load Actual DP across Pressure Temperature Enthalpy Error
primary superheater , [MPa] [°C] [kJ/kg] [kJ/kg] 40 % 0% of total 8.2 372 3055.2 6.2 85% of total 7.7 372 3067.5 -6.1 100 % 0% of total 18.2 401 2889.2 17.5 85% of total 17.1 401 2923.6 -16.9
Table 5.9 Extremes in conditions at first stage desuperheater inlet. 108
The maximum error possible by assuming that the primary and secondary superheaters carry an equal part of the pressure differential is 6.2 kJ/kg at 40% load and 17.5 kJ/kg at MI load. These errors represent 0.36% and 1.5% of the total heat rise in the first stage superheater at 40% and 100% load, respectively. (As the extremes mentioned above are mechanically not possible, the real error is much smaller than the 0.36% and 1.5%.) The error due to this assumption is therefore quite small, and the assumption holds true.
Based on the above argument, the assumption was also made that the any difference in pressure across the desuperheaters will have a negligible effect on steam enthalpy and may be ignored.
5.3.8 Spray Water Flow Measurement
The superheater spray water flow rate had an orifice and pressure differential measurement for the total spray water consumed by all four superheater desuperheater stations. The control system used a flow compensating function which took the square root of the pressure signal and converted the 0-100 % signal to a 0-70 kg/s signal. The reheater had a similar measurement for the total spray water flow to the two reheater desuperheater stations, which was ranged 0-25 kg/s.
Desuperheater spray flow calculations The spray water flow rate into a desuperheater was alsocalculated by means of energy and mass balance calculations across the desuperheater (Figure 5.7). The destIperheater inlet and outlet temperatures were measured at all the desuperheaters. Steain pressure was either measured, or calculated as described in the previous section. It is thus possible to calculate the enthalpy of the steam before and after all desuperheaters. The pressure and temperature of the spray water were also measured and its enthalpy calculated.
109
Spray water
m • h gPr
Steam in I Steam out Desuperheater m ; m; ho
Figure 5.7 Variables for heat balance calculations.
A list of the variables concerned is: desuperheater spray water flow rate
m0 = steam mass flow rate at desuperheater outlet ho = steam enthalpy at desuperheater outlet kor = spray water enthalpy h, = steam enthalpy at desuperheater inlet
m, = steam mass flow rate at desuperheater inlet
Equation (5.2) describes the conservation of energy across the desuperheater:
+ m /j (5.4) m0 ho = m1. hi spr spr
Equation (5.3) describes the conservation of mass across the desuperheater:
l = M. + M spr (5.5) o
For the superheater, all the enthalpy can be calculated, but only the superheater outlet steam flow is known. Thus:
(5.6) rn = m0 - map
Equations (5.2) and (5.4) may now be combined to form EqUation (5.5), relating desuperheater spray water flow rate desuperheater outlet flow. 110
mo (ho — h) ms pr (5.7) (h — h
Once the spray water flow rate was known for the second stage desuperheater, it was subtracted from the outlet steam flow rate to obtain the nett.inlet steam flow rate. The inlet steam flow rate equals the outlet steam flow rate for the first stage desuperheater. The superheater inlet flow rate was determined in the same way.
Unlike the superheater where the outlet steam flow was known, the reheater inlet steam flow was known. Equations (5.2) and (5.3) can be combined to form Equation (5.6); relating reheater spray water flow rate to enthalpy and the desuperheater inlet flow as follows:
. mi (h. — h) mspr (5.8) (ho — h sp)
The above methods were used to calculate the desuperheater spray water flow rates to all four superheater desuperheaters and the two reheater desuperheaters individually.
Spray water flow discrepancies It was noted that, on both superheater and reheater, discrepancies existed between the measured quantity of spray water and the quantity calculated by means of:a heat balance across the desuperheaters (Figures 5.8 and 5.9). It was important to identify the cause(s) of the discrepancies before deciding on which method to use as the best representation of the actual spray water flow rate. Three discrepancies were evident: a) The superheater spray water flow measurement was offset by 10 kg/s from zero. This flow discrepancy resulted from a permanent leakage of measured spray water to the economizer inlet. No isolating valves were present on the spray water warming lines, resulting in a continuous flow to the economizer (Figure 5.10). 1 1 1
(.7;80
at 70 ....0-... ° 60 50 -e- >,40 ..1-€ . 2- 30 4C-
92 20 « en 4 a)co 1 0 2 0 . • 0 10 20 30 40 50 60 70 Calculated spray water flow [kg/s]
Figure 5.8 Superheater spray water flow measurement.
7;30
c. x, 20 co >,
u) 10 E' sx co co a) 2 0 0 10 20 30 40 50 60 Calculated spray water flow [kg/s]
Figure 5.9 Reheater spray water flow measurement. 112
A To second stage desuperheater stations
A To first stage spray water nozzle
X Spray water hand isolating valves
el Spray water control valves
0-1 Spray water motorised isolating valves
Spray water warming lines Feed water line Spray water flow to economiser measurement Superheater spray water Boiler feed pump Feed water heaters
Feed flow measurement 'Feed regulating . valve
Figure 5.10 Superheater spray and warmup flow.
The leakage flow rate would have been dependent on the pressure differential between spray water and economizer inlet, which, in turn, depended on the differential pressure across the feed water regulating valve and the pressure loss in the spray water line. These variables vary somewhat during the regulation of drum level and steam temperature and will cause a slight variation in leakage flow rate. This variation in leakage flow rate could account for the larger variation on errors between measured and calculated spray water flows evident on the superheater as compared to the smaller variations evident on the reheater. The reheater desuperheater supply line had no warming lines connected to it. b) The calculation-to-measurement ratio was 0.55 for the superheater and 1.2 for the reheater. The ratio should ideally have been 1.00 for both. The error could have 113
resulted from thermocouple drift, flow measurement orifice calibration errors, or
in the case of the superheater (because the calculation showed less flow than measured) incomplete evaporation of the spray water at the position of the thermocouple pocket.
Thermocouple drift was subsequently tested on all desuperheater stations by comparing the desuperheater inlet and outlet temperatures after having shut off the
spray water supply to the respective desuperheaters. Differences between
desuperheater inlet and outlet temperatures were random and in the order of
0.5 °C. This could not account for the discrepancies observed between measured
and calculated spray water flow rates.
The completeness of spray water evaporation on the superheater was tested at
various spray flow rates by measuring the desuperheater outlet temperature at two
positions dovvnstream of the desuperheater, roughly one metre apart in the
direction of the steam flow. No significant temperature difference between the
upstream and downstream measurements could be detected. This indicates
complete evaporation of spray water before reaching the thermocouple pockets
(or, alternatively, a very good averaging of steam and water enthalpy by the
thermocouple pockets).
The calibration of the flow measurement orifices could not be verified because
design data was not available. However, because it was the only*other possibility
that remained, calibration errors on the desuperheater flow measurements were
assumed to be the cause of the calculation-to-measurement ratio not being unity.
c) The superheater spray water flow measurement saturated at 70 kg/s. This was as
a result of reaching the upper limit on the differential pressure transmitter.
Considering the above three points, determining the desuperheater spray water flow rate by means of heat balance calculations across the desuperheater appears to be a more 114
accurate method than the orifice and pressure differential method. Therefore, the spray water flow quantity used for modelling and control was calculated by means of a heat balance calculation across each of the desuperheaters.
5.3.9 Feed Water Flow Measurement As mentioned in the previous section, no isolating valves were installed on the spray water warming lines, resulting in a continuous flow to the economizer. This unmeasured quantity of water bypassed the feed water heaters and joined the feed water flow at the economizer inlet. This means that the quantity of water that passed through the feed water heaters was less than the feed water measurement. As a heat balance across the feed heaters was used to calculate the flow rate of the regeneration steam extracted from the cold reheat line, it was important that the feed water flow through these heaters was known reasonably accurate. Since the spray water line warming flow rate was not measured, it was estimated to be 10 kg/s, based on the offset of spray flow measurement above spray flow calculations. This flow is between 2 % and 5 % of the total feed water flow, so errors in this estimation would not have affected feed water heater heat balance calculations seriously.
5.3.10 Secondary Air Flow Measurement Based on the chemical composition of the coal burnt, a specific amount of air is needed for combustion. The ratio of air-to-fuel chemically required for combustion is called the stoichiometric air/fuel ratio. Kendal has a design stoichiometric air-fuel ratio of 6.34 kg air per kg fuel [36].
Under ideal conditions, stoichiometric air-fuel combustion will consume all the fuel and all the oxygen. In practice, boilers are run at a air-fuel ratio higher than stoichiometric to assist combustion efficiency [4]. Under these conditions the excess oxygen cannot be consumed and a certain percentage of oxygen will be present in the flue gas. Neglecting the minor effects of CO and NQ on the concentration of 0 2 in flue gas, it can be shown that the theoretical relationship between 0 2 concentration in flue gas and the ratio of stoichiometric air flow to actual air flow is described by Equation 5.9.
115
A s CO2F = 21 (1 - ) (5.9) AA
where: Co2F = Concentration of 0 2 in flue gas AA = Actual air flow rate A s = Stoichiometric air flow rate Equation (5.9) can be rewritten to obtain the ratio ofA s I A A in terms of Co2p.
A C S = 1 - 02F (5.10) A A 21 Both sides of Equation (5.10) may be inverted to obtain the ratio of actual air flow to stoichiometric air flow in terms of oxygen concentration in flue gas.
AA 1 (5.11) A s 1 - C02F 1 21 Large discrepancies were found between the calculated and measured ratios of A4 I As (Figure 5.11).
1.4
x x X x 4 • x x x54 a • a •x x se 1.3 x iee < To' x x E >cx 0 x 4c)a... E 1.25 YXX /le 'Sc xx x% O X co 1.2 X K )4 X X x x ,f, xr., X xt X„, T 0 .1. x :Et 1.15 , x le X X x X X XX
X XX 1.05 2 2.5 3 3.5 4 4.5 5 5.5 Oxygen in flue gas MI x Measured - Calculated
Figure 5.11 Discrepancies between calculated and measured air flow ratio. 116
Discrepancies between calculated and measured 0 2 in flue gas may be caused by any combination of: Inaccurate 02 concentration measurement The 02 concentration in flue gas was measured via two Zirconium-based sensors, one located on each side of the flue gas duct, above the air heater. Inaccurate fuel flow measurement Steady state mill fuel flow was measured based on the speeds of the two volumetric coal feeders located above each mill. Transient fuel flow measurement will be discussed later. Inaccurate air flow measurement The total air flow measurement comprised the sum of primary air flow and secondary air flow. Primary air flow was measured by means a duct venturi located upstream of each mill. Secondary air flow was derived from a differential pressure measurement on the inlet of each of the two secondary air fans and a precalibrated curve.
Accuracy of 02 measurement The accuracy of the 0 2 concentration measurement was tested and reported on earlier in the life of the power station [124]. The conclusion drawn in the report was that the Zirconium-based 02 measurement, as well as the plant installation was sufficient for accurate measurement.
Accuracy of fuel flow measurement Fuel flow and load generated are related through total plant efficiency, which varies only slightly through the operating envelope. A plot of generator load against fuel flow should therefore be a relatively linear curve. Assuming an accurate load transducer (which is a fair assumption since station revenue is based on this measurement), deviations from this linear curve indicate inaccurate fuel flow measurements. Figure 5.12 shows some discrepancies between fuel flow measurement and generated load, indicating an inaccurate fuel flow measurement at steady state.
117
700
§'600 2 Fo 500 O
O ro 400 C 6) 300
200 40 50 60 70 80 90 100 110 Fuel flow [%]
Figure 5.12 Correlation between fuel flow and generator load.
Errors in fuel flow measurement could be expected, since the coal throughput of a volumetric feeder is dependent on coal density. Fuel flow in Figure 5.12 actually refers to the energy input into the boiler which is also affected by the calorific value of the coal.
The Kendal boiler controls did have a long term correction, called fuel factor, that adjusted the fuel measurement so that the ratio of generator load to fuel measurement equated to 686 MW : 100 %. The fuel factor was automatically corrected by integrating the fuel flow excess / deficit with a time constant of 68 minutes and a dead band of about 32 MW. (The fuel factor adjustment was more complex than described here, but for the:purpose of this discussion, the above explanation is sufficient.) Variations in coal density and calorific value, combined with the slow correction rate and dead band, could have resulted in the deviations evident in Figure 5.12, but it needed to be established whether this was the cause of the discrepancies between measured and calculated AA / A s.
If the large discrepancies between measured and calculated air flow rates were caused by the erroneous fuel flow measurement, it should be possible to reduce the discrepancies if fuel flow is calculated from generator load rather than measured incorrectly from the plant. The fuel flow can be calculated based on the ratio of 686 MW = 100 % fuel. To test this, 118 measured and calculated AA ' A s were again plotted against 02 in flue gas, but this time the stoichiometric air flow rate A s was calculated using the fuel flow value derived from generator load (Figure 5.13).
1.4 X X K X • y X X .... ,-, X X • X X )..-. 0 • X X X X ._, X EP X X "%'. N.X SO°. X X Or X X X :<1 flee' X 5\ X X gek ,x . x _ -•37 !, x. . x. . >k * sr ' I :cT 1.15 - e. , x . . x
1.05 2 2.5 3 3.5 4 4.5 5 5.5 Oxygen in flue gas [%] x Measured .• Calculated
Figure 5.13 Air flow vs 02 in flue gas with fuel flow derived from generator load.
Large discrepancies between measured and calculated A4 I A s still existed, indicating that the errors did not originate with fuel flow measurement. Since 0 2 and fuel flow measurement errors have been ruled out to a large extent, the problem must be related to the air flow measurement.
Accuracy of air flow measurement Although there was no reference against which the air flow measurement could be verified, it was possible to compare the secondary air flow measurements made from the two secondary air fans. The two identical secondary air fans were constant speed fans with adjustable inlet vanes used for throttling air flow. The vanes on both fans were positioned to the same setpoint and the fans should therefore have delivered similar quantities of air. The identical instrumentation setup on the two fans should consequently have provided the same feedback signal of air flow.
119
To establish the integrity of the secondary air flow measurement, the difference in air flow
signals from the two fans were calculated for each heat distribution test. The result is
shown in Figure 5.14.
1 0
%) 5 [ ce I NIi 0 II III II I ICI 1 HI III 1 III 111111 1 III 1 I H11111111111111111 1 1 11 ' II
differen I 11111 IIINJIIIIIL 111111111i II 111111i1II 11.111111ill 111111111111 1 11111 HI t -5 1 I' I 1 1 1 III' I 11 1 1 1 1 11 1 -10 l' 11 1 III 1 15 1 1 ' Measu remen
20 Test #
Figure 5.14 Normalized difference between LH and RH air flow measurements.
Figure 5.14 shows large and erratic differences in air flow measurements taken from two
similar fans operating with similar inlet vane positions. Consequently, it was deemed to
be errors on air flow measurement that contribtited most toward the discrepancies noted
earlier between measured and calculated air flow rates. This was an important observation
which ruled out the use of air flow rate in the boiler modelling process.
5.3.11 Calorific Value of Coal
The calorific value of the coal burnt during the heat distribution tests were tested on a daily
basis. A maximum variation of 10% in calorific value was observed. The lowest tested
calorific value was 19.0 MJ/kg while the highest was 21.0 MJ/kg. No on-line
measurement of calorific value existed, so it was not possible to do direct compensation
on the mill fuel flow signals.
However, as described before, a compensator in the boiler controls adjusted a parameter
called the fuel factor. The fuel factor was adjusted according to any error between
estimated fuel requirement and actual fuel flow for achieving a certain load. The 120
assumption was made that all mills burned fuel with the same calorific value, so that the
fuel factor could be used to correct their respective fuel flows.
In practice, different mills could have burned coal with different calorific values, because
mill bunkers are filled one at a time, and not all at once. If the coal quality changes,
different bunkers will contain variations in coal quality while the fuel factor would have
been adjusted to the average coal quality. Unfortunately, no better way of compensation
existed.
5.4 Neural network model
Once the heat absorption has been calculated, the next step was to train an artificial neural network to model the heat transfer to the boiler elements, based on the test data obtained (Figure 5.15).
The model had to provide a nonlinear mapping between the furnace elements influencing the heat transfer rate and heat distribution, and the boiler elements to which the heat is transferred
(Equation 5.12).
= f (R) (5.12) where:
aff, = vector of modelled heat transfer rates to the boiler elements nonlinear mapping function
vector of furnace conditions affecting heat transfer rate
.Inputs from Modelled furnace heat transfer elements to boiler elements
Neural network furnace to boiler heat transfer model
Figure 5.15 Furnace to boiler heat transfer mapping. 121
Various aspects of neural network modelling as well as the training and testing procedure used during the modelling phase will be discussed next.
5.4.1 Training and testing data
From the 129 sets of valid test data, 116 sets were used as training data for the neural
network and 13 sets were used as test data. The purpose of the test data sets was to
determine how accurately the neural network represented data sets that it had not been
trained on. In other words, testing data was used to determine the generalization
capabilities of the neural network.
5.4.2 Training algorithm
The PC programme Brainmaker [83] was used to train the neural networks. The weights
in the network were initialised with pseudo-random values. Training was done by means
of the error backpropagation method. A training session typically consisted of 6000
training runs through the 116 training data sets. The neural network RMS output errors
on training data sets were recorded after each training run. Then the 13 sets of test data
were run through without training to obtain the RMS error on testing data.
The weights were automatically adjusted after completion of every training run.
Histograms of the weight values were then updated. The histograms were used as an
indication of the degree of saturation of the neural network (weights saturating at -8 or
+8), which in turn, indicates that the neural network is too small [83] (this reference refers
to a neural network with a high degree of saturation as being brain dead) During the
training session, the weights of the neural network were saved to a file every 50 training
runs so that an optimum set of weights could be selected afterwards.
5.4.3 Selecting the best network during training
The RMS error during the training phase and RMS error during the testing phase were
then plotted against number of training runs. While both errors decreased, the neural
network was constructively learning the input-output mapping. Should the error on
training data decrease while the error on testing data had increased, it suggests that the
122
neural network was learning specific data sets while its generalization capabilities were decreasing [83].
0.1
0.08
i. 0.06 a) Testing cn 2 0.04 Training
0.02
0
Training cycles (1-6000)
Figure 5.16 Error on test data increases after many training runs.
Figure 5.16 shows atypical training session. Initially, the training and testing errors are very large, but these reduce rapidly during the first few training runs The error on training data is (as expected) lower than the error on testing data. The averages of both errors decrease until, after many training runs, the error on testing data starts to increase due to the loss of generalization. A neural network selected from the area of minimum error on testing data was deemed to have the best input-output mapping (although this was not proved). Since the network weights were saved every 50 training runs, the network closest to the minimum RMS error on testing data was selected as the final result of the training session.
5.4.4 Evaluating the networks The steps described above were repeated three to four times on similar sized networks with weights initialized differently. The networks trained differently and had different best run errors. The set of neural network weights giving the lowest error was selected as the best possible with the specific architecture. 123
Once the "best" set of neural network weights was identified, it was loaded into the Brainmaker neural network software. The entire set of test data (129 points) was run through the network without training while the neural network outputs were written to a file and imported into a spreadsheet. There the neural network model outputs were compared to values measured on the plant during testing, and the error was established. In this way, the modelling errors from different networks could be compared on a test-by- test basis.
5.4.5 Network architecture and selection Thirty different feedforward neural network architectures were tested to obtain the optimum nonlinear mapping of the furnace input elements to the boiler output elements. As no firm network sizing theory has been established, experimentation with different neural network sizes was the best way to obtain the smallest neural network that still had good accuracy. The network sizes tested, ranged between zero and 160 hidden neurons in zero to three hidden layers. Every layer had one bias neuron of which the output was set to unity.
The following notation will be used to describe neural network topology: Input Neurons : 1st Hidden Neurons : 2nd Hidden Neurons • Output neurons. For example: 7:15:5:3, refers to a neural network with two hidden layers, 7 input neurons, 15 neurons in the first hidden layer, 5 neurons in the second hidden layer, and 3 output neurons. The bias neurons are not indicated, but it may be assumed every layer, except the
output later, has one bias neuron .
The neural networks that were tested had similar inputs, but their internal structures and outputs were different. The next few sections deal with the input, structure, and output of the various neural network models that were tested.
Inputs The input vector (u) to the heat transfer model comprised the furnace elements that affected heat transfer rate of heat distribution. These were the furnace elements 124
manipulated during the heat distribution tests. The same seven inputs were used during all tests and across all the network topologies. The input vector had the following manipulated furnace elements:
A-Mill - E-Mill fuel flow rate 0 % - 115 % 02 concentration in flue gas 2.3 % - 5.5 % Burner tilt angle -30° - 30°
The mill fuel flows were corrected by multiplying them with the fuel factor before being used as inputs. In this way, abetter representation of energy input could be obtained with changing coal calorific values. The concentration of 0 2 in flue gas was used as input in place of the total air flow measurement, because of the poor accuracy and repeatability of the latter. The average of all the burner tilt angles, as measured on the plant, was used for the burner tilt angle input.
Heat transfer to individual boiler elements The calculated heat transfer rates were available for each of ten individual boiler components. At first, a complex approach was followed whereby the model output vector (q,,,) comprised the heat transfer rates to each of the ten individual boiler elements. These were: Economizer Evaporator Left-hand primary superheater Left-hand secondary superheater Left-hand final superheater Right-hand primary superheater Right-hand secondary superheater Right-hand final superheater Left-hand reheater Right-hand reheater 125
Networks of various sizes were trained with this output configuration. The smallest RMS error on test data for some network sizes are tabulated below. Although the average RMS error over all outputs was not excessively high, large, unrepeatable errors were evident when comparing the individual model outputs to calculated heat transfer rates. In some cases, errors in superheater component model outputs were as large as 50%. However, when the sum of the heat transfer rates to all superheater components were compared to the sum of these model outputs, the errors were more acceptable.
Network architecture Lowest RMS error on testing data 7:21:10 5.13% 7:50:10 4.75% 7:28:36:10 4.45% Table 5.10 Results of networks trained with 10 individual outputs
The reason for the poor modelling accuracy could be that the individual boiler components receive different air flow streams that vary in velocity and temperature (especially on the superheater) and the size of the components are not large enough to represent an average heat transfer. Small changes in burner tilt angle disturb these flow patterns and thereby have a large, almost random effect on heat transfer to the individual components.
This was observed in practice too, where the right-hand side of the reheater requires more desuperheating than the left-hand side for certain burner tilt angles, and less for other angles. If these large differences in heat transfer are still present in the back-end of the boiler, it must also be present at the furnace outlet where most of the superheater heat transfer takes place.
As the intended use of the neural network model was for the control of heat transfer to the superheater and reheater, it was not necessary to model the heat transfer to every individual component. The heat transfer to the superheater as a whole and reheater as a whole was of prime importance. Due to the high errors, and no real need for the ten individual heat transfer outputs, this model was not developed any further. 126
Heat transfer to grouped boiler elements The objective of the heat distribution controller was to control the heat transfer to the reheater and superheater. The minimum outputs required from the heat transfer model are therefore heat transfer to the superheater and heat transfer to the reheater. Based on the minimum requirements and on the large errors obtained with the previous model, it was decided to group the heat transfer rates to all the individual superheater components into one variable and similarly, heat transfer to the reheater components into one variable. For the sake of completeness and ease of error detection, heat transfer rates to the economizer and evaporator were also grouped to obtain one variable.
The new neural network model still had the same input vector described previously, but the output vector (g,,,) had only thiee outputs namely: Evaporator heat transfer rate Superheater heat transfer rate Reheater heat transfer rate
Networks of various sizes with the grouped outputs were trained on the same training and testing data as before, using the same procedure. Best run RMS errors for different network sizes are given below in Table 5.11. " The increase in error with the very large network is probably due to a decrease in the generalization ability of the network.
Network size Lowest RMS error on testing data [%] 7:15:3 5.17 % 7:50:3 4.05 % 7:14:6:3 4.70 % Table 5.11 Results of networks trained with 3 grouped outputs
Although the RMS errors of the grouped output models were on average only slightly smaller that of the individual output models (Table 5.12), the errors between the model outputs and calculated plant heat transfer rates were more acceptable. For example, no test conditions on the 7:50:3 network resulted in model output errors greater than 20 %. 127
Output Characteristic Size Lowest RMS test Error [%] Individual components 7:50:10 4.750 Grouped components 7:50:3 4.054 Table 5.12 Comparison of individual to grouped output heat transfer model.
The model output errors for all 129 tests are shown in Figure 5.17 for the best 7:50:3 network. Modelling errors on the reheater output are noticeably larger than that of the evaporator and superheater. Similar observation were also made in practice during reheater steam temperature controller tuning. The reheater seemed to "act differently" from day to day and between consecutive tests.
0.3
0.2
Evaporator
Superheater
Reheater 0.2
0.3 Tests 1 to 129
Figure 5.17 7:50:3 neural network model output errors for all tests.
Three separate networks Since the modelling ability of the neural network seemed to improve somewhat when the complexity of the output pattern was reduced, a test was devised to establish the ability of a neural network to model only one specific boiler component. The output training data was split into three sections, one for heat transfer to each of the three grouped components, i.e. evaporator, superheater, and reheater. Three neural networks were trained individually on the three sets of data. The three 7:5:1 networks each had one hidden layer with five 128
hidden neurons to compare the results with that from the one 7:15:3 network obtained
previously. No major difference in accuracy was noted between the single network model
and the three network model (Table 5.13).
Error [%]
Output Characteristic Size Evaporator Superheater Reheater Average Grouped components 7:15:3 3.00 3.70 8.81 5.17 Three networks 7:5:1 (3) 2.88 4.08 8.80 5.26 Table 5.13 Comparison of two output strategies.
Relative heat transfer to grouped boiler elements
Heat transfer rate to the individual boiler components is quite linear relative to boiler fuel
input. As boiler load increases from 40 % to 100% the heat transfer rates to all boiler
components change through a factor of about 1.5 while the heat transfer rate to any
component in relation to the others changes only slightly. By analysing the heat
distribution test data, it was established that the ratio of heat transfer to the evaporator,
superheater, and reheater is normally close to 50:30:20.
When changing boiler load, variations in absolute heat transfer are far greater than
variations in relative heat transfer. Inaccuracies in a model of absolute heat transfer (as
done up to now), may then overshadow the subtle changes in relative heat transfer. Since
the total boiler heat transfer rate is proportional to fuel flow, this need not be modelled.
What needs to be modelled are the changes in heat transfer of the individual components,
relative to total heat transfer. In this way the model will be trained on variations in heat
distribution which can be superimposed on the linear (relative to fuel flow) heat transfer
rate.
This scheme was tested by training a neural network model on outputs expressed as a ratio
of total heat transfer. 129
Y m= f(u) (5.13) where: r,„ vector of modelled relative heat transfer ratios
Once the network was trained, the model outputs were multiplied by boiler efficiency and total heat discharge rate (derived from fuel flow rate and the heating value of the fuel) to obtain the absolute heat transfer rate to the individual boiler components.
(5.14) gm 'C m q qf where: qf total furnace heat discharge 77 boiler thermal efficiency
Errors between modelled and actual heat transfer rates were much lower with the relative heat transfer model than with the absolute heat transfer model (Table 5.14).
Error [°/0]
Output Characteristic Size Evaporator Superheater Reheater Average Absolute heat transfer 7:15:3 3.00 3.70 8.81 5.17 Relative heat transfer 7:15:3 2.33 2.20 5.46 3.33 Table 5.14 Improvement in results by modelling relative heat transfer.
Corrected heat transfer Since the neural network model was trained on heat transfer rates relative to the total heat transfer rate, the sum of the neural network outputs should ideally be unity. This was not the case in reality, where the sum of the model outputs was close to, but usually not equal to unity. Varying with different models and model inputs, the sum of the model outputs ranged between 0.97 and 1.03. 130
Because the sum of the model outputs should be 1.00, it should be possible to correct any
deviations from unity by proportionally adjusting the model outputs. This was done by
setting the corrected model outputs equal to the individual model outputs divided by the
sum of the model outputs.
an, (5.15) am, = rm
where:
effIC vector of corrected modelled heat transfer rates r„, = scalar sum of relative heat transfer rates
Since this correction was used to adjust outputs of a trained neural network, it did not
affect the training of the networks. The same networks that had been trained previously
on relative heat transfer rates could have the output correction done. Table 5.15 indicates
the improvement in accuracy achieved with output correction.
Error [%]
Output Characteristic Size Evaporator Superheater Reheater Average
Relative heat transfer 7:15:3 2.33 7 2.20 5.46 3.33
Relative + correction 7:15:3 1.86 1.96 4.64 2.82
Table 5.15 Improvement of accuracy by correcting the outputs.
With the heuristically motivated adjustments made to the heat transfer model, it was
possible to decrease the average RMS error from 5.3 % to 2.8 % for similar sized
networks. The heat transfer model with corrected relative heat transfer outputs was the
most accurate configuration achieved with the neural network model, and for this reason,
it was the configuration used in the heat distribution controller.
Figures 5.18, 5.19, and 5.20 compare output errors across the 129 data sets for the last
three neural network configurations shown in Table 5.16. 131
Output Characteristic Size RMS Error on test data [%] Individual components 7:21:10 5.13 Three networks 7:5:1 (3) 5.26 Grouped components 7:15:3 5.17 Relative heat transfer 7:15:3 3.33 Relative + correction 7:15:3 2.82 Table 5.16 Comparison of different heat transfer model results
Determining the network size Once the final neural network model output configuration had been established, different network sizes were tested to find the smallest network with a good representation of the heat transfer. As before, four training exercises starting with different randomised initial weights were done on every selected size. The 7:15:3 network had the best accuracy. Network models with more and less than 15 hidden neurons displayed a reduction in accuracy. The main results are shown in Table 5.17.
Error [%]
Output Characteristic Size Evaporator Superheater Reheater Average Relative + correction 7:30:15:3 1.971 1.938 4.654 2.85 Relative + correction 7:15:10:3 1.846 2.058 4.662.= 2.86 Relative + correction 7:15:3 1.861 1.961 4.641' 2.82 Relative + correction 7:10:3 2.519 2.709 5.380 3.54 Relative + correction 7:5:3 2.424 2.326 5.591 3.45 Table 5.17 Summary of results obtained from different network sizes
132
0.3
0.2
Evaporator
Superheater E 8-0.1 z Reheater -0.2
-0.3 Tests 1 to 129
Figure 5.18 Absolute heat transfer rate model.
0.3
0.2
Evaporator
SS 0 To Superheater
Reheater -0.2
-0.3
Tests 1 to 129
Figure 5.19 Relative heat transfer rate model errors.
0.3
0.2
Evaporator
Superheater
Reheater -0.2
-0.3 Tests 1 to 129
Figure 5.20 Corrected relative heat transfer rate model errors. 133
Mapping discrepancies One point of concern toward the final stages of selecting an optimum network was discrepancies in the input-output mapping of identical networks with different initializations. This became apparent only after the spreadsheet model was available and testing the networks was simpler.
To demonstrate the discrepancies, three sets of weights were obtained for three 7:15:3 networks initialized differently. These weights were loaded into the software model (Appendix C) where five similar input scenarios were entered and the outputs noted. The input scenarios were: burner tilt angle at 0 0, 02 at 3.5 %, all mills in service with four at 50 % load and the other mill at 100%. Each of the five scenarios had a different mill loaded to 100%. Irregular and sometimes large discrepancies in modelled heat transfer rates were observed (Table 5.18).
Mill loaded to 100% Heat to Evaporator Heat to Superheater Heat to Reheater
A Weights set 1 694 516 288 Weights set 2 702 526 303 Weights set 3 688 536 273
B Weights set 1 709 503 286 Weights set 2 835 - 473 224 Weights set 3 677 553 267
C Weights set 1 731 504 262 Weights set 2 737 528 268 Weights set 3 717 510 270
D Weights set 1 736 481 ,/ 281 Weights set 2 738 527 vz 267 Weights set 3 719 500 277
E Weights set 1 742 491 265 Weights set 2 749 527 256 Weights set 3 735 496 266
Table 5.18 Heat transfer rates obtained with different initializations.
After this observation was made, many more 7:15:3 neural networks were initialised randomly and trained in the same fashion. One neural network was then selected on the basis of an average representation of heat transfer rates. 134
5.4.6 Comments on accuracy
The errors obtained with the neural network model were of similar magnitude as errors obtained through a multi-input multi-output boiler process model based on state-space equations with parameters derived via autoregression techniques documented by Aitchison e.a. [39]. Apart from the difference in modelling techniques, the model described here is a nonlinear heat transfer model designed for the entire boiler load range as opposed to the linear model designed by Aitchison e.a. [39] which was essentially an interpolation between three models derived at three different operating points.
Although the neural network model needs less prior plant information than analytical or state-space techniques, the process of obtaining the "best" model proved to be quite time consuming. It may be remarked that much of the process of training, evaluating, and selecting an optimum neural network could be automated by means of a computer programme. 135
6. Neural networks and steam temperature control •
6.1 Requirements for improved steam temperature control To improve the quality of steam temperature control on power plant boilers above that possible with normal HD feedback control, the causes of bad or sub optimal control need to be addressed. The factors leading to poor steam temperature control were discussed in an earlier section. Briefly, these are: disturbances (especially load ramps and mill changes / trips), long time lags, process parameters changing with time and load (time constants, heat transfer rate, steam properties), dead time, control loop interaction, and over-firing. The design of an advanced steam temperature controller should therefore be geared towards addressing as many as possible of the factors leading to poor steam temperature control. The desired controller characteristics to fulfil this design are discussed below.
6.1.1 Predictive control When properly tuned, conventional feedback control can regulate steam temperatures adequately under steady state conditions. This is an idealistic case since various process disturbances will affect the steam temperature. Examples of these disturbances are: fuel type, burner tilt angle, excess air, blowdown, steam bleed, load ramps and coal mill changes / trips. Some of the disturbances may affect the steam temperature much quicker than the feedback control is able to respond, causing temperature excursions. The most severe steam temperature excursions originate from disturbances in load and firing system (see Page 48). Both of these are measurable disturbances. If the effect of the disturbances on heat transfer or on steam temperature can be predicted in advance, appropriate control actions can be calculated and executed with minimal disturbance on steam temperature.
An advanced steam temperature controller should have an appropriate process model that can predict the effects of a disturbance. The controller must also have some algorithm to calculate the appropriate control action for cancelling out the effect of the disturbances. Ideally a process model should be used to predict the effect of the control actions too so that the controller can balance the effect of the control action to that of the disturbance.
136
Figure 6.1 shows the broad structure of a model-based predictive controller. Two models predict the effects of the disturbances and control actions, respectively. A comparator feeds the difference between the two predictions to a controller, that calculates a control action which aims to balance the control and disturbance effects. As a secondary regulation action, any differences between the real process variables and their respective setpoints are fed back to the controller.
Disturbances Control Disturb. ---jai•• effecti-74(-- : effect 4 model model Set Points —00 + _A
Controller Process
Control signals Regulated process variables
Figure 6.1 Model - based predictive control.
6.1.2 Nonlinear control Due to the nonlinear behaviour of the boiler and steam generating protess, optimum control response can not be achieved across the entire operating range with linear process prediction models and a linear controller. This places severe restrictions on the use of classical control theory, based on linear differential equations.
The nonlinear modelling and control capabilities of neural networks have already been motivated. Based on these capabilities and on the documented successes with neural networks in nonlinear control applications, it seems feasible to employ neural network technology for creating the process model and controllers for steam temperature control. 137
The main requirement for using neural network technology (other than an appropriate controller structure) is that the network must be trained on masses of data. This data is already available during normal running of the power plant but can also be acquired during special tests (as done in this case). An advantage with doing special tests is that, when properly planned, many plant and process characteristics may be extracted over a relatively short time duration.
6.1.3 Adaptive control Changes in boiler parameters due to boiler tube sooting, changes in coal properties, variations in feed water temperature etc. necessitates that the steam temperature controller is adaptable to sustain optimum control. The mechanism of adaptation is to compare the output of a process model or a controller to some desired output. Adjustments are then made to some parameters internal to the model or controller to drive the difference to zero (Figure 6.2).
In•uts Model or Actual output
Desired out uts
Figure 6.2 Adaptive adjustment concept.
6.1.4 Heat distribution control For any arbitrary steam flow rate, some design rates of heat transfer to the superheater and reheater exist (Figure 6.3). The design heat transfer rates are adequate to raise the enthalpy of the steam and obtain the desired outlet steam temperatures. Deviations from
138
design heat transfer rates would cause temperature deviations had it not been for the closed loop automatic control system keeping steam temperatures at setpoint. Since the rate of closed loop control action is dictated by the long process time constants (5 - 10 minutes), the closed loop correction is quite slow.
800
2 600 iL5
ait 400 C 200 a)
---- 0 0 100 200 300 400 500 600 Steam flow rate [kg/s] — Evaporator --- Superheater Reheater
Figure 6.3 Design heat transfer rates to maintain steam temperatures.
Large disturbances can occur on the fire side of the boiler. Disturbances like load ramps and mill trips were shown to cause substantial temperature excursions due to a large and almost instant change in the distribution of the heat discharge. These rapid changes in heat distribution are the cause of steam temperature excursions.
Since the major disturbances all occur on the fire-side, i.e. rapid changes in mill firing rate, it would be beneficial to eliminate it at the source. It may not be possible to prevent mill trips and load changes, but it may be possible to maintain a constant heat flow rate to the superheater and reheater. Many control elements exist in the furnace for manipulating heat distribution. These are. the individual mill firing rates, furnace air flow rate, and burner tilt angle. Maintaining design heat transfer rates to the superheater and reheater will improve steam temperature regulation.
From a temperature control perspective, it is not necessary to maintain design heat flow 139
to the evaporator because disturbances there, will not directly affect steam temperatures.
During load up-ramps an excess in heat flow exists due to over-firing. It is desirable to direct the excess heat to the evaporator to assist the boiling process. In doing so, heat
transfer to the superheater and reheater can be kept to design (based on steam flow) to prevent temperature increases. During down-ramps a deficit in heat flow rate exists due
to under-firing. Then heat must be directed away from the evaporator to the superheater
and reheater in order to maintain steam temperatures. Directing the heat away from the
evaporator will also reduce boiling and assist in decreasing the steam flow rate.
6.2 Optimal heat distribution control
Based on the reasoning presented in the previous subsection, the author proposed a scheme in which the available fire-side control elements are manipulated in such a way that heat is distributed optimally between the different boiler components. The heat requirements of the different boiler components will be calculated on-line, and the furnace conditions will be adjusted to meet these requirements. Under conditions where the heat distribution cannot be made equal to design, the excess or deficit in heat transfer to the superheater and reheater will be calculated and the desuperheater spray water flow rates will be adjusted accordingly. This new control scheme will be referred to as Optimal Heat Distribution (OHD) control.
6.2.1 Available fire-side control elements
Due to the large spacing between the Kendal burners, the bottom mills are more suited to
producing pressure, and the top mills are more suited to producing teniperature. The
bottom mills discharge most of their heat onto the water walls of the boiler, thereby
producing steam flow, while heat discharged from the top mills is mostly superheating the
steam. The burner tilt angle has a similar effect on heat distribution. It is therefore
possible to alter the distribution of heat between the evaporator and the superheaters by
biassing the mills in service and by altering the burner tilt angle. On the other hand, an
increase in furnace air flow rate leads to increased convective heat transfer. By
manipulating the 02 setpoint, it is possible to alter the distribution of heat between the
superheater and reheater due to the superheater having both radiant and convective surface
but the reheater having mainly convective surface. Therefore, by biassing the firing rate 140
between mills, changing the burner tilt angle, and manipulating furnace air flow, the heat distribution between the evaporator, superheater and reheater can be influenced.
6.2.2 Controlling heat distribution The standard boiler controls are configured so that all mills in service are fired equal, burner tilt angle is determined according to mill combination (see page 37), furnace air flow demand is calculated from fuel flow rate and 0 2 setpoint, and the latter basically follows a predefined curve with some correction for reheater temperature condition. The OHD controller was designed to intercept these control signals, predict the resultant heat distribution, compare the distribution with design values, correct the control signals if necessary, and pass them on to the cascade controllers (Figure 6.4). Should the available control elements not allow total correction of heat transfer, the OHD design allowed for the utilisation of feedforward signals to the desuperheater controllers to do the necessary preventative adjustments to the spray water flow rate.
Existing Boiler feedback steam Desuperheater heated temperature control signals and control control elements
Desuperheater feedforward signals Individual mill fuel demands Fuel demand )• Existing boiler Furnace 0, set point Optimal heat pressure, air flow, control distribution Air flow set point and burner tilt controller elements controls Burner tilt set point Burner tilt set point
Figure 6.4 Signal flow to and from the optimal heat distribution controller.
6.2.3 Advanced feedforward with original feedback The OHD controller was conceived to be an advanced feedforward calculator to identify and counteract disturbances on the fire-side of the boiler. The original feedback steam •
141
temperature controllers would therefore remain active for normal temperature regulation. However, should a mill trip or a load ramp start, the furnace elements will be manipulated by the OHD controller to maintain the design heat transfer rate to the superheater and reheater. The OHD action was designed to be an open loop controller. Model inaccuracies and unmodelled disturbances leading to steam temperature deviations will be trimmed out by the normal closed loop steam temperature controls.
6.3 Controller design As motivated in Section 6.1, the OHD controller should possess predictive, nonlinear, and adaptive properties, and it should optimize heat distribution to balance out fire-side disturbances. The design of the OHD controller was done to incorporate these requirements. The main aspects of the controller design are discussed in this section.
6.3.1 Heat transfer error prediction
Heat transfer model Requirements of predicting the heat transfer rate via a nonlinear model were satisfied by using the neural network model trained on real boiler data. Inputs to the neural network were conditions on the furnace side (mill firing rates, 0 2 measurement, and burner tilt angle) and outputs were the predicted heat transfer rate to the boiler components (evaporator, superheater, and reheater). The predicted heat transfer rates were obtained via the neural network as functions of the furnace conditions:
qep = fe(furnace conditions) (6.1)
grafi = gfurnace conditions) (6.2)
qrp = fr(furnace conditions) (6.3)
where: predicted heat discharge to evaporator
gsp predicted heat discharge to superheater predicted heat discharge to reheater and:
fe = neural network mapping of evaporator 142
= neural network mapping of superheater
f, = neural network mapping of reheater.
Design heat transfer
Design heat transfer rates based on steam flow rate were calculated at increments of 50
kg/s between 200 kg/s and 600 kg/s from test data. The neural network heat transfer
model running on a spreadsheet described previously was used to calculate these design heat transfer rates. A balanced boiler was assumed with A, B, D, & E-Mills in service and the burner tilt angle set to 0°. The 0, input was kept to the design curve, based on steam flow. Mill demands necessary to obtain the different desired steam flow rates were calculated and entered into the model inputs. The model outputs under the various conditions were recorded to be used as design values. A look-up table with interpolation was used to obtain continuous smooth heat transfer rates as a function of main steam flow rate:
(6.4) qed = fed( nms)
(6.5) qsd = fsd( n ins)
(6.6) qrd = frd(n..) where:
qed design heat discharge to evaporator
qsd design heat discharge to superheater
qd design heat discharge to reheater and = fed design heat transfer curve of evaporator
fed = design heat transfer curve of superheater
frd = design heat transfer curve of reheater rnms = main steam flow rate.
Effect of disturbance
The predicted and design heat transfer rates were compared to obtain the predicted effect of a disturbance. Figure 6.5 shows the basic configuration of the nonlinear error predictor. 143
Any disturbance in the furnace conditions shows up as an error in heat transfer rate.
ee = qed qq) (6.7)
(6.8) es qsd gsp
(6.9) er qrd qrp where:
ee heat discharge error to evaporator
es = heat discharge error to superheater
er heat discharge error to reheater.
Mill firing rates Evaporator error ). )10 0 2 measurement +A - Su erheater error
Burner tilt angle eheater erroio_
Neural network heat transfer model
Steam flow rats
Design heat transfer curves
Figure 6.5 Predictive calculation for error in heat transfer.
6.3.2 Heat transfer optimization Once the predicted heat transfer errors are available, the new control action must be calculated. The backpropagation technique was already motivated as a practical and easily applied way to obtain the derivatives of the error on the inputs of the model. With a neural network as the process controller, the error derivatives can be backpropagated through the controller network and its weights adjusted accordingly. In the case of OHD control where a feedforward action must be calculated to counteract fire-side disturbances, the 144 error derivatives may be used directly to adjust the control elements.
Problem statement A vector of fire-side control signals must be obtained, which will minimize an index of temperature excursions, J. Also called a cost function, J, must be chosen in such a way that, through its minimization, the errors between the design heat transfer rates and the predicted heat transfer rates will also be minimized. A convenient choice of J is the sum of the square of the errors in heat transfer rate.
1 J = — (ti e 2 + ases2 + at?) (6.10) 2 ee where: ae = gain factor on the evaporator heat transfer error as = gain factor on the superheater heat transfer error a, = gain factor on the reheater heat transfer error
The gain factors allow for changes to be made in the relative importance of errors on different components. For example, by setting a, = 0, heat transfer errors to the evaporator will be ignored. Equation 6.10 can be minimized by backpropagating the errors through the heat transfer model and adjusting the control elements in the direction of the partial derivatives [76].
Backpropagation Derivatives Errors firin g r ates_ Evaporator
.4 S2 meas urement Superheater -4 Burner tilt angle Reheater Neural network heat transfer model
Figure 6.6 Backpropagation of errors to obtain derivatives.
An iterative optimization routine was designed to perform the following steps: 145
Calculate the design heat transfer rate to all boiler elements, based on the main
steam flow and a set of design curves.
Set biassing on all OHD control elements to zero. This removes OHD control
alterations and restores the existing boiler control signals. (The control signals are not passed on to the plant yet) .
Calculate the heat transfer rate based on the boiler control signals plus biasses by
means of the neural network heat transfer model.
Calculate the errors in heat transfer and adjust these through multiplication by the
respective gain factors (a, = 0, a , = 5, a,. = 5).
If the sum of the adjusted errors is less than a predefined limit (3Mi/s), go to Step
10.
If the decrease in sum of adjusted errors from the previous iteration is less than a
predefined limit (1.5 AN), go to Step 10.
Backpropagate the adjusted errors through the heat transfer model to obtain the
error derivatives with respect to the network inputs (control elements).
Add the error derivatives to the respective control element biasses.
If the number of iterations through this routine exceeds a predefined limit (50), go
to Step 10.
Go to Step 3.
Output the boiler control signals plus biasses to the control elements or secondary
controllers.
This routine minimizes the heat transfer error by adjusting the control elements based on the backpropagation algorithm. Whenever a control element is placed in manual control, its bias is reset to zero before every iteration and the feedback signal (mill fuel flow, measured 02, etc) is used as input to the neural network model.
Running on a 100 MHZ Pentium PC, one thousand iterations, each consisting of the feedforward neural network calculation, the backpropagation routine, and control signal adjustments, could be executed in 220 ms with a neural network size of 7:15:3. 146
With the gain factors set to 5 (as indicated above), the optimization procedure converged within 25 to 30 iterations. Figure 6.7 shows the bias development during the iterations of an optimization run of a simulated load ramp. Figure 6.8 shows the consequent reduction in errors.
40 30 20 a 10 Cw 0 cti di -10 -20 -30 -40
Iterations (1 to 25)
— A Mill— 02 — C Mill— D-Mill— E Mill— Tilts
Figure 6.7 Bias development during an optimization run. Note that B-Mill is out of service.
800 123 700 7600 – :7500 2400 – 03 "a 300 –
I0 200 100 Iterations (1 to 25)
— Evaporator — Superheater— Reheater
Figure 6.8 Heat transfer errors during an optimization run.
A disadvantage of using the backpropagation of error method is that, since the algorithm 147
is based on gradient descent and because the error surface may possess multiple local
minima, the. optimization routine may converge to a solution which is locally, but not globally optimal [114]. This did happen in practice, and will be discussed later. It was also said earlier that the heat transfer model has no defined inverse because many input conditions may exist for the same output condition. This argument still holds, but since the control element biasses are all set to zero when the optimization routine starts, the algorithm will converge to the solution closest to the initial conditions. This is desirable, because in practical terms this means that the solution requiring the least biassing will be obtained. It is desirable to bias power plant control elements as little as possible to reduce wear and tear on the plant and minimize maintenance costs.
6.3.3 Optimising the fuel flow rate
During the start of a load ramp, excess fuel enters the boiler by means of over-firing. The heat transfer model indicates excess heat transfer and the heat distribution optimizer quite simply biasses the mills downward (as they were before the ramp). This action eliminates the error in heat transfer rate, but it also eliminates the over-firing and therefore eliminates the load ramp. To prevent the above from happening, a fuel flow optimization routine was designed to achieve three things:
Ensure that the sum of the individual mill fuel demands meets the total fuel flow
demand.
Ensure that the individual mill fuel flow limits are adhered to.
Bias the mills as close as possible to the demand of the heat distribution optimizer.
The fuel flow rate optimisation was done iteratively by executing the following steps:
Add the mill bias signals to the original (unbiassed) mill demands.
Adjust all biasses that cause mill demands to exceed upper or lower limits.
If number of iterations exceed a predefined value (100), go to Step 9.
Calculate sum of biassed mills demands.
Subtract sum of biassed mill demands from total fuel demand to obtain file! error.
If fuel error is smaller than a predefined margin (0.1%), go to Step 9.
Add fuel error to all mill biasses. 148
Go to Step 1. Output biassed mill demands to the individual mill fuel controllers. Fuel flow rate of any mills running in manual mode were taken into account by using the mill fuel flow feedback signal, and no bias adjustments were made to the setpoints of these mills.
If the total fuel demand exceeded the capacity of all mills, Step 6 would never be true and the algorithm would never terminate, therefore the inclusion of Step 3. This modification was made after a live test during which OHD control was shut down by the unit controls while doing a downward ramp at low load. The unit pressure controller requested less fuel than achievable with all mills at minimum fuel demand. The fuel flow optimizer then continued looping through, because the total fuel demand and individual mill demands could not be matched. While stuck in this loop, a watchdog timer built into the boiler controls timed out, and OHD control was shut down.
This algorithm normally converged in 15 to 23 iterations and took less than I ms to complete.
6.3.4 Calculating desuperheater spray flows A very powerful advantage of having a boiler model is that a fairly accurate estimate of the heat surplus or deficit to the boiler elements becomes known. This enables the new control system to balance out disturbances in heat transfer by adjusting the amount of desuperheating on the reheater and superheater without having to wait for temperature changes. Thus, apart from reducing the disturbance in heat transfer by manipulating furnace elements, OHD control was also designed to calculate the exact amount of desuperheating spray water needed to maintain steam temperatures during transients.
The calculation of the amount of spray water can be demonstrated by the following example. Consider an upward ramp in boiler load. Assume that, after biassing the control elements to their limits, an excess of heat transfer to the reheater is still predicted. Reheater spray water.flow rate needs to be increased to prevent a temperature deviation. 149
Additional spray water must be injected so that, when this spray water is evaporated and
heated to the design reheater outlet temperature, it had consumed exactly as much heat as
the predicted excess heat transferred. The additional spray water mass flow calculation
is given by:
9 ex I77 — (6.11) spr h — h rh s spr
where:
mspr = spray water mass flow
e/ex = excess heat transfer hth, = outlet enthalpy of reheat steam
h spr = enthalpy of spray water
This spray water demand was divided by the number of desuperheater stations on the
boiler element (four on the superheater and two on the reheater), and the final value
obtained was the amount of change spray water required from each desuperheater to balance out the error in heat transfer to the boiler element. During a downward load ramp a deficit in heat transfer may occur. Then the spray water mass flow is a negative value.
This is quite achievable, because under steady state conditions the steam temperature controller is already injecting spray water. The spray water flow rate will then be reduced by the amount calculated above.
Unfortunately the desuperheater cascade slave controllers are not mass flow controllers, so the required spray water mass flow rate cannot be requested directly. Instead, they work as desuperheater outlet temperature controllers (see Page 56). The setpoint to the desuperheater slave controllers are made up of the output of the master controller (the master being the final steam temperature controller), plus a feedforward bias (Page 58).
It is this feedforward that OHD control was designed to manipulate. Consequently, the spray water flow rate bias derived above had to be converted to an outlet temperature bias.
To do the conversion between spray water flow and temperature, a heat balance calculation was done across the desuperheater. The outlet enthalpy was obtained as
150
follows:
117/7i ho - (6.12) tn +mspr where: m, = steam flow rate into desuperheater h, = enthalpy of steam at desuperheater inlet ho = enthalpy of steam at desuperheater outlet
Once the outlet enthalpy had been calculated, the outlet temperature was obtained from on-line steam tables. The feedforward signal was set equal to the desuperheater inlet temperature minus the desuperheater outlet temperature.
Therefore, during an upward ramp in load, the desuperheater slave controller receives a decrease in setpoint as a result of the OHD feedforward signal. The slave controller responds by opening the spray water control valve to inject more spray water in order to match the desuperheater outlet temperature to the lower setpoint. Once the outlet temperature matches the setpoint, the additional spray water injected is just enough to absorb the excess heat transfer caused by the over-firing.
6.3.5 Adaptation Adapting the design heat transfer curves Consider a condition where the boiler is running with the heat transfer rates at the design point, without any biassing from OHD control. Assume that the lowest mill in service is shut down for maintenance. Due to the mill being shut down, the natural heat distribution pattern inside the furnace changes. To keep the heat distribution on the design curves, OHD control increases the firing rate on the lower mills, reduces the firing rate to the upper mills, biasses the tilt angle down and possibly changes the furnace air flow rate too. As long as the mill in question remains out of service, the furnace will be operating in this biassed condition.
Operating a biassed furnace is undesirable from an operating and OHD control point-of- 151 view. Firstly, operators prefer the mills to fire at equal rates, because they use the mill fuel flow indications for early warning signs of milling problems. Also, when inspecting furnace flame formation, it is difficult to identify an out-of-normal flame if the mills are not fired equally. Secondly, for OHD control to reject disturbances, the control elements need room to move. When the control elements are biassed due to mill combination, the capability of eliminating disturbances are reduced in one direction. In the above example the fireball is already biassed downward to compensate for the bottom mill which is out of service. If an upward ramp in load is done, OHD will try to lower the fireball to prevent overheating of the superheater. With the plant already biassed, there may not be enough control action left to prevent temperature excursions.
Consequently, a long-term correction must be done to the design heat transfer curves to relax the OHD control action and reduce biassing. This was done by adjusting the heat transfer curves according to the real heat transfer rates, calculated from plant measurements. The adjustments were done by multiplying the design heat transfer rates of the individual boiler components with adjustable correction factors. The correction factors were in essence the integrals of the difference between corrected design heat transfer rates and actual heat transfer rates (Figure 6.9). It can be argued that the correction must be done far slower than the longest process time constants. By trial and error, the design correction time constant were set to 2400 seconds (40 minutes). Once the design curves have been multiplied by this correction factor, they are referred to as the heat transfer target curves.
However, when a disturbance on the furnace-side occurs and OHD control compensates totally by means of biassing the control elements, no error will remain and the design curves will not be updated. For this reason, the OHD action had to be stopped prior to achieving total disturbance rejection. At first, a variable was introduced to reduce the biassing by a certain percentage (10 % worked well) after the final bias calculation by the optimizer. For example, if the tilt angle was to be biassed by 20°, the bias would be reduced by 10 % and the tilts would only be biassed by 18°. This resulted in the error required between design and actual heat transfer, which forced an adjustment of the design 152
curves.
Correction factors
Design heat Steam Target heat transfer flow rate transfer rates calculation
Inputs from plant Actual heat absorption calculation
Figure 6.9 Adjusting the design heat transfer to match plant conditions.
Although the desired effect was achieved by this reduction in biassing action, another method was later applied. The new method applied the full biassing to the control elements, but it placed a dead band of 2.5 MJ/s on the individual heat transfer errors before calculating the bias. By adding a dead band on the error signals, the errors are effectively reduced in magnitude before being received by the optimizer. After optimising the heat distribution, a small error, unknown to the optimizer, still remains between the design curves and the actual plant condition. This error forces the adjustment of the design curves as described above.
The dead band method was preferred over reducing the bias because it also acted as a filter for small variations in heat transfer around the design points under steady-state conditions. In both cases, slight temperature deviations were expected as a result of limiting the OHD action, but this would be taken care of by the normal closed loop controls. At full load, a sustained 2.5 MJ/s error on heat transfer to the superheater will cause a temperature deviation of only 1.6 °C. As the design curves are adjusted to reduce the errors, the reduction in errors results in a reduction in biassing, which in turn sustains the errors. This process repeats until the design curves represent the new state of the process without any 153 biassing.
Adapting the heat transfer model Many factors can influence the rate of heat transfer to boiler components (see Section 3.2). Some of these factors (such as burner tilt angle and air flow rate) are measurable and can be modelled, but others (such as boiler sooting and seized burner tilt mechanisms) cannot be measured easily and are therefore not modelled. Unmodelled process characteristics lead to an inaccurate process representation, and to prevent erroneous heat transfer predictions, the process model must be adapted or recalibrated on-line.
Adapting a neural network model can be done in two ways: Retraining the network on new process data This method has the advantage that any change in furnace characteristics will be modelled. For example, if baffles are installed in the furnace to reduce the flue gas velocity past tube banks susceptible to ash erosion, these changes will be captured in the model. However, the model may be totally corrupted by a faulty sensor reading burner tilt position. Also, the training data must be sufficiently rich with heat transfer characteristics to obtain a representative model. But in practice, the plant may be run around full load for extended periods of time. If model adaptation is needed, it will have to be done at the full load condition, and training on the full load data will cause the model to "forget" the low load data, resulting in an inaccurate low load model. On-line model correction through training could work well under "laboratory" conditions, but it needs careful consideration before applied in practice. This method was not used to adapt the OHD control model.
A linear correction of model outputs. Even if the neural network model is not retrained on new data, it is still possible to do model correction. The model outputs are simply multiplied by a correction factor, similar to the method described for design curve correction. Errors between the corrected model outputs and real plant heat transfer rates are used to adjust the model output correction (Figure 6.10). The correction time was determined by trial and error and finally set to 900 seconds (15 minutes). This linear correction method has also been used successfully for 154
correcting the outputs of a state-space, linear regression, boiler model [52].
Correction factors Furnace conditions Corrected heat transfer predictions
Inputs from plant Actual heat absorption calculator
Figure 6.10 Adjusting the heat transfer model to match plant conditions.
6.3.6 On-line steam tables Many calculations relied on the availability of the enthalpy of water or steam. To obtain the enthalpy on-line, a special algorithm was developed for the calculation of steam and water properties. As for the steam properties calculations used during the modelling phase, the calculations were based on the IFC formulations of the thermodynamic properties of water for industrial use [122]. This method differs from the lookup table method generally applied [39] & [54].
Although the IFC formulations are very complex and part of the calculations use iterative algorithms, the entire set of enthalpy calculations for the boiler model (26 points) executes in only 50 ms on a 100 MHz Pentium PC.
6.3.7 Main control algorithm The main control algorithm was timer driven with a cycle time of one second. The following functions were performed during each cycle: Read inputs from plant Calculate heat absorption 155
Calculate design heat transfer
Calculate predicted heat transfer
Calculate errors and optimise heat transfer through backpropagation
Write outputs to plant
Adapt design heat transfer calculation
Adapt heat predicted heat transfer calculation
Update graphics and write variables to file
The control algorithm executed within 500 ms on an Intel Pentium running at 100 MHz.
6.4 Expected results
With the spreadsheet neural network heat transfer model and the standard built-in Quattro Pro optimizer, an OHD control emulator was built. The example of the mill trip used in Chapter 3 was optimised by the OHD emulator to demonstrate the expected improvements in boiler heat transfer obtainable though OHD control. Table 6.1 shows the expected improvements in the heat transients occurring during a mill trip. Large improvements are shown for both the superheater and reheater in a trade-off with the less sensitive evaporator.
Boiler element Existing Heat Optimal Heat Relative Transient Transient Improvement
Evaporator - 6 MJ/s + 8 MJ/s - 33%
Superheater + 48 MJ/s + 7 MJ/s 85%
Reheater - 42 MJ/s - 15 MJ/s 64%
Table 6.1 Improvements in heat transfer after a mill trip
The improvements in heat distribution are achieved by manipulating the furnace elements
to minimize the difference in heat transfer before and after the mill trip. Table 6.2 shows
how the furnace elements are set up for optimal heat distribution. 156
Furnace element Existing control OHD control
A-Mill demand 93.3% 62%
B-Mill demand 93.3% 115% C-Mill demand 0% 0%
D-Mill demand 0% 0% E-Mill demand 93.3% 110%
02 Setpoint 3% 5. 5%
Burner tilt angle -15° -30°
Table 6.2 Furnace element setup after a mill trip.
The optimal heat distribution controller will take similar measures to optimize heat transfer during load ramps. 157
7. Practical implementation and results
7.1 The PC as control platform
7.1.1 Kendal boiler control system The boilers at Kendal Power Station are controlled via the ABB Procontrol P13 distributed control system [125]. This system is not flexible enough to accommodate advanced control schemes such as Optimal Heat Distribution control. For example, the 70PR03 control processors can do only integer arithmetic. Its programmable memory is limited to 64 kB and the programme resides in an EPROM.
7.1.2 System requirements for advanced control Advanced control schemes as the one described in this thesis are best developed and tested on one of the modern versatile platforms (like Unix or Windows), using a flexible programming language (like C, Pascal or Visual Basic). Therefore, prototypes of advanced control schemes are in most cases not done on the existing plant control system, but on a programmable personal computer that can do floating point number calculations and has a large memory area. For example, the advanced boiler control strategy developed in Microsoft C by Hitz e.a. [54] used a 386/387 industrial PC running MS DOS, since the process computer could not support the large volume of floating point calculations required nor had it storage space for steam tables. March [52] states data logging, colour displays, and flexibility as the motivations for using a PC for modelling and control of steam temperature on a nuclear plant.
7.1.3 OHD control hardware The Optimal Heat Distribution control scheme would perform an enormous amount of data processing due to the neural network model, the optimisation routines and the on-line steam table calculations. Because the system would be used on-line for real-time control, the computer had to have ample processing capacity. With these processing requirements in mind, a 100 MHz Pentium computer was used for control. The computer was located in an air-conditioned and vibration free control room, so it was not necessary that the 158
computer be an industrial computer. The control computer was equipped with 16MB RAM, a CD ROM drive for loading software, a 1.44 MB stiffy drive for downloading data, keyboard, mouse, and a video display adapter for presenting the graphic screens.
7.1.4 OHD control software Initially, the operating system of choice was Windows NT [126], due to it being a proven 32-bit multitasking system. However, after comparing this package to Windows 95 [127] on a cost-benefit basis, the latter took preference. Because networking or multi-tasking was not envisioned for the control computer, no good reason could be found for running the control software on the more expensive Windows NT package. The only problem experienced with Windows 95 was that all application executions are paused while a window is being resized or dragged across the screen. This problem was later solved by loading Microsoft PLUS! [128], which allows background processing while windows are dragged or resized. The programming language used for writing the control algorithms was C++ [121] and the graphics were done with a charting tools package [129].
7.1.5 Operator / Engineering interface The PC screen, keyboard and mouse were used as an operator / engineering interface. The PC screen was a 17" super VGA monitor for large, clear display in the control room. Two charts were displayed on the screen. The first was a set of design heat transfer curves for the evaporator, superheater and reheater, also indicating the actual heat transfer and the predicted heat transfer with the control elements biassed and unbiassed. The second was a bar chart indicating the unbiassed control element demands from P13, the degree of biassing done by the OHD controller, and the control element feedback signals from the plant. A screen dump of the graphic display is shown in Appendix D.
Engineering access was provided to display and change the internal OHD control parameters. The OHD control programme could also be executed from within the Borland C++ integrated development environment in a debugging mode which gave access to all the programme variables, and enabled the execution of the algorithms to be traced. Both these facilities proved very useful for programme maintenance, debugging, and OHD 159
control optimisation.
7.2 Interfacing to existing boiler controls
7.2.1 Communications hardware All the control modules in the ABB Procontrol P13 distributed control system are interconnected through an ABB P42 Intraplant Bus system [125]. Because all the data needed from the plant can be made already available on any the P13 local control busses, the most cost effective method of data acquisition was to read the required signals directly off this system. This was done via a ABB 70BK03 bus coupler. This device is an RS485- to-P13 bus interface. The RS485 serial output of the bus coupler was connected to an RS485 serial interface card on the OHD computer.
As the 70BK03 and the RS485 interface supports bi-directional communication, the same hardware used for reading inputs from the plant can be used to write the control signals from the computer back to the P13 control system. A diagrammatic layout of the interface between the 01-ID controller and the P13 system is shown in Figure 7.1.
Pentium PC for P13 Boiler Control System OHD control
r —
Hardwire Data Link II I 'gat< 70 BK 03 In I RS 485 Bus coupler Interface Card
Figure 7.1 Interface between PC and existing boiler control system.
7.2.2 Communications protocol The ABB system has a proprietary serial communications protocol. A communications module was programmed as part of the OHD control programme to request all the 160
necessary data points from the P13 system and store these in allocated variables for use by
the control programme. Once the control task has been completed, the control signals
were written back to the P13 system, where the -original control system executed the control requests.
Internal P13 variables are 16 bits wide (or a word) and represent numbers scaled between -200% and 199.97% in 0.024% resolution. In hexadecimal format the word may range between 0000 and FFFF. The serial communications protocol sends these data words coded in hexadecimal format by using ASCII characters. The communications module in the OHD software converted these ASCII characters to a 4-byte string which was then converted from hexadecimal format to a fraction of unity represented by a floating-point number. All variables were then converted to the appropriate engineering units.
The maximum speed of the communications link was 38400 bits per second and used 10 bits to transmit a byte. OHD control read in 64 data values @ 10 bytes/value and wrote out 15 values @ 14 bytes/value. Assuming negligible processing time, the communications part of the programme took 220 ms to execute. Time consumed by the communications routine alone was about as much as the rest of the entire programme, graphic displays included.
7.2.3 Fail - safe operation
OHD control was designed to run in parallel with the existing boiler control system so that it could be shut down at any time without detrimental effects on the boiler. This was a requirement for fail-safe implementation and for doing alterations to the system with the boiler on load. It also made possible a comparative evaluation with the advanced control turned on and turned off. This approach was also followed by others [39], [55], and [64].
The OHD programme could also be run in Standby mode in which all the control modules were being executed, but the biassed control signals were not sent back to the P13 system.
The original control signals were just mirrored back to the P13 system when OHD was in standby mode. 161
OHD control was turned ON and OFF from the operator control panel. When active, the OHD control programme generated a 0.5 Hz binary square wave signal which indicated to the boiler controls that the OHD computer is functional, that the OHD programme is being executed and that the OHD control mode is Active : Should no transition on this signal be present for three seconds, the P13 control system switched out the OHD control. The OHD control signals were stored inside the P13 system on the BK03 bus coupler, so that in the case of the OHD computer failing totally, the last control signals still remained active until the OHD control was switched out of the control circuits.
7.2.4 Closed loop controls OHD control was not designed to perform any closed loop control. All the normal closed loop controllers in the P13 system remained active regardless of the state of the ORD system. However, three control signals were 'intercepted' by the 01-1D control system and modified before being routed back to the P13 system.
Signal switch selector P13 OHD P13
OHD control selector [ ÷÷. A-Mill fuel control Optimal heat B-Mill fuel control Analog signals )0 ,.. distribution e lqtre control C-Mill fuel control Binary signals algorithm D-Mill fuel control
Boiler pressure control • )1. E-Mill fuel control
02 set point generator 02 control
Tilt set point generator Tilt positioners
Figure 7.2 Closed loop control signal flow diagram.
One of the signals routed through the OHD computer, the total fuel demand signal from the boiler pressure controller, was split into five individual mill control signals which could be modified individually before being routed back to the mill fuel controllers. Analog 162
signal switches were programmed in the P13 system through which the source of the control signals could be selected. Figure 7.2 shows the signal flow routes between the P13 and OHD systems.
7.2.5 Feedforward temperature controls OHD control was configured to take over the existing feedforward control signals to modify the desuperheater outlet temperature setpoints. A similar approach is also described in [45]. Irrespective of the feedforward signals generated by the P13 system, OHD calculated new feedforwards and sent these back to the P13. system. Ana—log signal switches were programmed in the P13 system through which the source of the feedforward signals could be selected. Figure 7.3 shows the feedforward signal flows between the P13 and OHD systems. The feedforward input signals to OHD were used purely for
comparison purposes when 01-11D was active and were mirrored•to the outputs if OHD was in standby mode.
P13 Signal switch selector OHD OHO control selector
Analog signals OHD Binary signals control algorithm P13
1st stage Shtr 1st stage Shtr
2nd stage Shtr 2nd stage Sit
LH Reheater LH Reheater
RH Reheater RH Reheater
Figure 7.3 Feedforward control signal flow diagram.
7.2.6 Fault tolerance The P13 control system was provided with interlocks so that OHD control could only be selected to operate if all its input signals were available and within a realistic range. This 163 method was also used by Aitchison e.a. [39]. For example, the OHD control mode could not be switched on unless both HP feed water heaters were in service. This was a requirement for proper reheat steam extraction calculations.
On the other hand, the OHD system performed many internal checks before writing back new control signals and changing the state of the 0.5 Hz signal. One of the obvious checks was to see if the boiler is operating in an area contained in the training data of the model i.e. steam flow rate between 200 and 600 kg/s. Other important checks were also done, such as ensuring that the enthalpy of steam is used (and not that of water) at the desuperheater outlet under saturated conditions. (this check was built in after a major calculation error occurred when the enthalpy of water was returned by the enthalpy calculator. The calculation was correct, the plant measurements not.) With Windows as an operating system, it is possible to simultaneously run multiple instances of the one application. This is undesirable for a control application, and a feature was built into the OHD programme to prevent the execution of more than one instance of the programme.
7.2.7 Commissioning the system The new software for the P13 system was loaded and a BK03 bus coupler was installed on Kendal Unit 3 during an outage. At this time, the OHD control programme was still under development. Since the P13 serial communication protocol is ASCII text-based, the serial interface between the OHD computer and the P13 system could be tested using the Windows 95 Hyper Terminal software. The new P13 software was cold=commissioned using simulation modules to generate and check test signals. The power plant was returned to service normally.
The OHD control programme was developed off-line and tested on simulated data. After connecting the OHD computer to the P13 system via the serial communications link, the software communications module was tested and the programme was run in standby mode while executing the control algorithms using real plant data. Some minor programming errors were corrected during this period. Once the control programme worked satisfactorily, the control output signals from the ODD computer to the P13 boiler control 164
system were next in line to be commissioned. Since it was the first time that the OHD generated signals would be actively used for control, clearance for a 'Risk of Trip' was obtained from the national load control centre.
Eleven control signals were read in from the P13 system and duplicate signals were sent back as control signals. These were: 1 - 4) 4 * feedforward signals to spray water flow controllers. 5 & 6) Tilt position setpoint and 0 2 setpoint. 7 - 11) A-Mill to E-Mill demand signal.
The necessary diagnostic hardware was coupled to the P13 system and the eleven control signals were commissioned one-by-one through the next three steps: Ensure that the ABB P13 system is receiving the correct value on the signal. Toggle the software switch inside the P13 system via a simulation to activate the signal. Monitor that the ABB P13 system responds correctly.
The 0.5 Hz binary signal, the calculated enthalpy of main steam and reheater spray flow rate signals were also sent from OHD to P13. These signals were commissioned at the same time as the eleven control signals.
Apart for some minor problems with the communications software module, the signals were commissioned as planned. Once all the signals were checked and activated, all the simulations were removed to restore the signal flow paths to normal. The OHD control program was modified to do zero biassing and the system was turned ON and OFF from the control room. Since the P13 - OHD interface was designed to be fail-safe, this too was tested by activating the OHD control system and, while active, the OHD computer was turned off. The boiler controls recognised the failure and switched back to normal control mode without incident. The OHD control system and the P13 interface was then declared ready to run the advanced control software. 165
7.3 Steady state testing and optimization
Initially, OHD control was turned on during steady state conditions. Two problems
appeared which had to be rectified before transient testing could commence. The first problem occurred as a result of mills running with an offset on fuel flow, and the second
problem was due to process variation.
7.3.1 Mill fuel offset
The fuel flow rate of the Kendal mills is raised by increasing the primary air flow to the mill
and reducing the mill bypass damper position to force more primary air through the mill
(Figure 7.4). The primary air flow is in direct proportion to mill fuel demand and the
bypass damper position is based on mill demand and a precalibrated curve, called a mill load line.
Air / Fuel to boiler
Air + fuel out
Bypass damper -4On Nunn NE
Air in
Coal Mill • Primary air .
Figure 7.4 Mill bypass damper and air flow paths.
Apart from a limited degree of correction done automatically on the bypass damper
position, the mill fuel control is essentially an open loop control system. Should the ball
charge of a mill run low, less fuel is produced with constant primary air flow and bypass
damper position. Consequently, the mill fuel flow falls beloW the demand and an
uncorrected offset on fuel flow develops. A mill is then referred to as 'running off its load
line'. Due to the open loop control the offset between mill demand and actual fuel flow 166 remains until the ball charge is replenished.
Since there are no serious operating consequences to a mill running off its load line (except at very low loads, where the mill fuel flow may decrease below the trip value), mills are often run for days with this offset between mill demand and fuel flow. However, the underproduction of the mill alters the furnace heat distribution pattern slightly. The target heat transfer rates are updated by the OHD controller to reflect this altered heat distribution. When the optimisation routine is run, it recognises that this one mill has to under-produce to match the target heat distribution. Therefore, it biasses the mill down below setpoint. On receiving this reduced setpoint, the mill controller reduces the primary air flow to the mill and opens the bypass damper, which reduces the fuel flow rate from the mill even more. Again, the heat distribution is altered, the target curves are adjusted and the mill setpoint is reduced even further by the optimizer. The scenario escalates until the mill demand is blocked by the lower limit.
To prevent this escalation, mills running with an offset in fuel flow had to be compensated for. This was done via a fuel error estimator, which adjusted a variable called the mill fuel error over a period of time (Figure 7.5). The time constant of the correction was determined by trial and error and set to 450 seconds (7.5 minutes).
Mill fuel error ■•■111.0. Mill fueldemand
Mill fuel measurement ON- +
Figure 7.5 Error estimation on mill fuel flow.
Should a mill be under-producing, the mill fuel error had been added to the unbiassed mill demand signal before the latter was sent to the optimizer, so that the optimizer used the 167
correct fuel flow rate when predicting heat transfer rates. Based on the corrected fuel flow rates, the heat transfer predictions matched the target heat transfer rates and no further biassing was required.
7.3.2 Process variation Most of the measured signals indicated a certain degree of variation in the process variable. These variations are natural for the process and originate from small disturbances and control actions. For example, all the Kendal units are utilised for power regulation. As the power demand on the national grid varies with loads being switched in and out, the generator loads are automatically increased and decreased by a few MW from the national load control centre. This results in the fuel flow rate to the boiler varying almost continuously. Even these small variations in fuel flow resulted in small errors between predicted and target heat transfer rates and subsequent biassing of control elements. Although the variations in unit load cause variations in steam temperature, most of the variations are small (see Page 49) and do not cause concern. The control element biassing performed by the OHD control system was deemed unnecessary and had to be inhibited to reduce wear and tear.
To inhibit the unnecessary biassing of control elements, a dead band was placed on the error between target heat transfer and predicted heat transfer (see Page 152). The dead band was set to eliminate all errors smaller than 2.5 MJ/s. The control actions were also affected by measurement noise. First order lags were added to the bias path of the control outputs to smooth down the operation. It is important to note here that the base control signals as generated by the P13 system were not filtered to prevent inducing additional phase lag into the system. Only the bias values were filtered. The filter time constants were set to 10 seconds.
7.4 Transient testing and optimization Once the steady state performance of the ODD control was improved, transient tests were done. Most of these tests consisted of load ramps, but mill trips, and a load runback test were also done. All the tests were done twice, once with only the normal boiler controls active as a reference, and 168 then with 01-1D control active. Before discussing the final results, various problems that were experienced will be discussed and their respective solutions presented.
7.4.1 Undesirable optimization The first test was a load ramp from 686 MW to 586 MW at 15 MW/min with B, C, D, & E mills in service. The biassing worked as expected for under-firing, the upper mill and burner tilt was biassed upwards to make up for the loss of heat to the superheater and reheater (Figures 7.6, 7.7, and 7.8). However, a glitch occurred in the mill and tilt biasses (Figure 7.7 and 7.8) and the biassing seemed to disappear for a while. This was not expected, since neither the fuel flow rate or steam flow rate displayed an uneven gradient.
105
100
95
7 so
85
80
75 — Fuel flow [%] Steam flow index
Figure 7.6 Fuel and steam flow rates during a down ramp under OHD control. (Time over X-axis = 30 minutes)
It was suspected that this undesirable biassing action was caused by the neural network and backpropagation optimizer converging into local minima with sub-optimal heat distribution results. The recorded data was run through the optimizer again off-line and it was confirmed that convergence into a local minimum caused the incorrect biassing. By altering the recorded data, it was established that other minima existed too. Adding a momentum term to the gradient decent was tried, but did not improve the situation. To overcome the local minima the momentum term had to be made so large that it frequently caused instability during convergence. 169
30
20
10
Figure 7.7 Burner tilt angle during load ramp, showing optimization glitch.
110
100
90
*--.0 80
7 70 --ran- - -"m- 60 IN basee IIIr -Mr 50 11/ 40 — B-Mill — C Mill — D-Mill — E Mill
Figure 7.8 Mill demands during ramp, showing biassing error.
Different network sizes were then tested. The larger networks were found to be more prone to local minima than smaller networks. This observation makes sense from a curve- fitting perspective. As a simple case, with three coordinates on an x-y plane, the quadratic function y = ax2 + bx + c can be determined unambiguously. If a higher-order curve is fitted to the same three data points, many fits are possible, and local minima could be 170
created (see Figure 7.9). A reduction in polynomial order may be thought of as the curve being stretched tighter between points, consequently reducing the formation of unwanted minima.
Figure 7.9 Different polynomials fitted to the same three points.
It was therefore strived to find the smallest network size that still provided fair modelling accuracy, to reduce the occurrence of local minima. The same training and selection procedure described in Chapter 5 was used. Results on accuracy obtained with various network sizes are presented in Table 7.1.
Based on the increase in error obtained with networks containing less than 5 hidden neurons, it was decided to change the 7:15:3 heat transfer model with the 7:5:3 one. This decision was based on a trade-off between a reduction in model accuracy and the aim of reducing localized minima. Although the numerical values show a 9 % increase in error due to reducing the number of hidden neurons from 15 to 5, a graphical comparison of the errors between modelled and actual heat transfer over the 129 tests, shows that no serious reduction in quality was induced (Figures 7.10 and 7.11).
▪
171
Network size Overall RMS error rk] 7:15:3 2.82 7:10:3 3.54 7:7:3 3.14 7:5:3 3.08 7:4:3 3.87 7:3:3 4.12 Table 7.1 Accuracy of networks with various numbers of hidden neurons.
0.3
0.2
0.1 Evaporator ..1LiAtAti• .0 0 '01Tryr ' Superheater E 5-01 Reheater -0.2
-0.3
Tests 1 to 129
Figure 7.10 Modelling errors with the 7:15:3 network.
0.3
0.2 6 0.1 Evaporator
.) .0 0 v 1•••••1 Superheater r E r Reheater -0.2
-0.3 Tests 1 to 129
Figure 7.11 Modelling errors with the 7:5:3 network. 172
The 7:5:3 neural network was then loaded into the model and this network configuration was used for all the following tests.
7.4.2 Cycling During the transient tests, the fuel flow tended to oscillate when 01-ID control was active. Figure 7.12 shows this cycling as recorded during a load ramp from 686 MW to 586 MW at 15 MW/min with A, B, D, & E mills in service. The oscillations caused the control elements to be biassed in an oscillatory fashion (Figure 7.13 and 7.14).
105
100
.9)- 95 In2 8 so
80 Time (30 minutes)
Figure 7.12 Oscillating fuel flow during down ramp under OHD control.
Fuel flow rate is the manipulated variable for boiler pressure control. Increased firing increases steam production, but steam flow to the turbine is kept constant by the generator load controller through throttling down the governor valves. The excess steam production therefore increases boiler pressure. The pressure controller is therefore tuned based on the pressure response of the boiler in relation to fuel flow changes. The pressure controller settings are calculated based on the pressure response obtained when fuel flow is directed through all mills simultaneously, and without any burner tilt movement. 173
15
0
5
Time (30 minutes)
Figure 7.13 Burner tilt action to regulate heat transfer to superheater and reheater.
110
100
90
80 pip 70.
60
50
40
Time (20) minutes — A-Mill B-Mill D Mill — E Mill
Figure 7.14 Mill biassing to regulate heat distribution.
When OHD control is active, an increase in fuel flow is directed mainly through the lower mills (the upper mills may even reduce their fuel flow) while the burner tilt angles are decreased. These actions are aimed at directing the additional heat away from the superheater. The excess heat is then directed towards the evaporator where it augments the boiling process. When boiler load is decreased, the opposite happened.
Because the excess / deficit heat discharge is directed to the evaporator under OHD 174 control, the boiler steam production response in relation to fuel flow differs from the normal response from which the pressure controller settings were calculated. Due to the evaporator receiving much more fuel during an up ramp and much less during a down ramp, the gain of the fuel-to-pressure process is increased by OHD control. This was tested in practice by making step changes in total boiler fuel flow with a constant generator load setpoint, first with normal boiler controls (no biassing) and then with OHD control active. The results in Figure 7.15 show the faster boiler pressure response when OHD control is active.
tL 0 0 0
z ...... 0 0 0 1 1111141111111111111114111111
Time (15 minutes)
— Pressure Fuel flow
Figure 7.15 Boiler pressure response to fuel flow with OHD control on and off.
From a practical perspective, when the steam pressure is slightly high, the boiler pressure controller decreases the fuel flow rate. Predicting the deficit in heat transfer to the superheater, OHD control tilts the burners upward and increase firing rate on the upper mills while reducing the firing rate on the lower mills. While this action is beneficial for the heat transfer to the superheater, the evaporator loses much more heat than the pressure controller anticipated. This causes the pressure to decrease faster than expected and the pressure controller is caught off guard. By the time the pressure controller responds, the boiler pressure has decreased significantly, and a large quantity of additional fuel is injected to reverse the pressure decay. With this large increase in fuel flow, OHD control predicts overheating of the superheater. Consequently the tilt angle is decreased, the lower mills 175 are fired harder and the process reverses. Continuous cycling results. This is all due to the process responding more than what the pressure controller was tuned for.
Final calculations showed a 30 % increase in process gain with OHD control active. New boiler pressure controller settings were then calculated based on the faster boiler response under OHD control. These settings were entered into the controller, but it was found that the pressure controller response became sub-optimal with less biassing. Depending on the degree of biassing of the control elements, the heat shift may be more, or less than obtained during the above test. The 30 % increase in process gain observed during the test will therefore not always be constant. An assumption of an average increase in process gain of 20 % was made, and new controller settings were calculated. These settings had to be entered manually each time before OHD control is turned on. Toggling between two sets of controller parameters can easily be automated, but in the case of the Kendal boiler controls, this can only be done during an off-load period.
Although the reduced controller gain did improve fuel-pressure cycling to a certain extent, under high degrees of biassing, the process gain was still increased significantly, and the cycling re-appeared. Large process disturbances, like mill trips and capability load runbacks, still caused process cycling (these test results will be shown later).
7.4.3 Fuel flow measurement errors
One of the observations made during up-ramp tests, was that the superheater temperatures decrease substantially under OHD control. Figure 7.16 shows the results of a load ramp test from 586 MW to 686 MW at 15 MW/min with A, B, C, & D mills in service. Load up-ramps under standard boiler control normally had the steam temperatures increasing due to over-firing. Under OHD control, the steam temperatures decreased despite the predicted heat transfer to the superheater, due to the biassing action, closely matching the target (Figure 7.17).
176
542 105
—540 a) 2.538 95 a) tca
2 536 90 E TD (1) 534 85
532 80
Time (30 minutes) — Main steam temp Fuel flow rate
Figure 7.16 Main steam temperature decreasing during load ramp under OHD control.
560
540
520 2 500 '17) • 7,,( 480
TO 460
440
420
Time (30 minutes) Predicted — Target
Figure 7.17 Predicted and target heat transfer rates to superheater during load ramp.
It was later established that, during an up ramp, a large discrepancy existed between the predicted heat transfer rate and the actual heat transfer rate calculated from plant measurements (Figure 7.18). The predicted heat transfer rate (or rate of heat discharge) matches the actual heat transfer rate (or rate of heat absorption) at the start of the ramp and shows a maximum deviation shortly after the end of the ramp. The deviation then
177 slowly decreases over an 8 minute period so that the two signals match again.
1600
.11500 0
@ 1400
1300 Time (20 minutes) — Absorbed Discharged
Figure 7.18 Discharged and absorbed heat flows.
The rate of heat absorption is calculated from plant measurements (see Page 95) and is believed to be an accurate representation of the true heat absorption. The total heat discharge is calculated from the fuel flow rate, the calorific value of fuel, and the boiler efficiency. The calorific value of fuel, and the boiler efficiency will not change sufficiently to cause deviations to the extent shown in Figure 7.18. This indicates an untrue fuel flow measurement during transient conditions.
The mill fuel flow measurement is actually a calculation, taking primary air flow and bypass damper position into account. The speed of the volumetric coal feeders is used as a long term correction on the fuel flow calculation, but during transient conditions, the fuel flow is derived only from the estimated air flow rate through the mill.
The dynamic response of a coal mill is discussed in depth by Peet e. a. [130]. On increasing the air flow rate through the mill, there is an initial proportional increase in mill coal output rate due to .the additional pulverized coal picked up by the increased air flow. The increased output eventually decays back to the original mill coal output rate since there is no corresponding increase in coal input to make up the coal deficiency in the mill (Figure 7.19). 178
70
65
0 6 T.) u_ 55
50
Time Figure 7.19 Mill fuel flow response to increased air through-flow. [130]
On increasing the coal input to the mill by increasing the coal feeder speed, there is a lagged increase in coal storage and coal output rate, provided that the mill is not flooded with coal. After a period of time determined by the mill system design, the coal output rate will settle out at a new value which matches the coal input rate (Figure 7.20).
75
70
E 65 0 a, 60 u_ 55
50
Time Figure 7.20 Mill fuel flow response to increased coal input. [130]
Coal mill controls increase the feeder speed and mill air flow rate simultaneously. The nett result is an initial quick increase in coal output rate followed by a drop in coal output and a second gradual rise to the steady state (Figure7.21).
179
75
70
65
0 r„ 60 LL 55
50
Time Figure 7.21 Mill fuel flow response to increased coal and air flow. [130]
Since the mill fuel flow measurement at Kendal does not take the above considerations into account, it is quite possible that the discrepancy between discharged and absorbed heat transfer during transients arise from the unmeasured and unmodelled mill dynamics.
This was verified by tripping one mill during four-mill operation while the unit maintains constant load. The three mills remaining in service were automatically ramped up by 30 % each, to maintain the total fuel requirement. Had the true fuel flow from these mills increased by 30 % each, no additional correction would have been needed. However, due to the mill dynamics described above, the mills did not produce the additional 30 % fuel each, and the total fuel demand was increased by the pressure controller to maintain steady unit load (Figure 7.22).
During the entire time span covered by Figure 7.22, the generator load and steam flow were constant. Therefore, the real fuel flow had to be reasonably constant. As a result of quick increase in fuel demand imposed on the three mills remaining after the trip, the mills indicated a higher fuel flow rate than actually produced. This is the same fuel flow measurement used by the OHD controller to predict the heat transfer to each of the boiler components, therefore the incorrect heat distribution.
180
75 Mill trip ;39 1-c— 70 0 iu c 65 0 60
LcCI H 55
Time (30 minutes)
Figure 7.22 Fuel flow indication increasing after mill trip.
Although generator load, or even measured heat absorption, could provide a more accurate total fuel flow indication during transient conditions, the OHD controller needs the fuel flow rate from each individual mill to calculate heat distribution. To provide the OHD controller with a better representation of actual fuel flow, a lead-lag compensator plus 3rd order filter was placed on the individual mill fuel feedback signals to mimic the 'mill dynamics (Figure 7.23). The time constants for the compensator were derived by trial- and-error.
Measured 1.25 s + 1 1.25 s + 1 Corrected fuel flow fuel flow
1 -3/11.1 (1.8 s + 1) 3
Figure 7.23 Correction circuit for mill fuel flow. Time constants are in minutes.
Heat discharge rate from the same load ramp as discussed earlier, was recalculated using adjusted mill fuel flow signals. The results are presented in Figure 7.24. Although not a perfect match, the heat discharge calculated from the adjusted fuel flow signal runs closer to the heat absorbed curve than the heat discharge calculated directly from the measured fuel flow signal. 181
1600
a 1500 0
1) 1400
1300
Time (20 minutes) — Absorbed Discharged — Adjusted
Figure 7.24 Heat discharge calculated from the adjusted fuel flow measurement.
Based on the improvement it brings to the heat transfer calculations, the adjustment to mill fuel flow feedback signals were implemented into the OHD controller.
7.4.4 The 02 control problem The error in fuel flow measurement did not only have an effect on the predicted heat transfer. Furnace air flow was also affected. The setpoint to the furnace air flow controller is calculated from fuel flow and the output of the 0 2 controller (Figure 7.25). When the fiiel flow rate increases, the air flow rate is increased proportionally, and with the output of the 0 2 controller increasing, a proportional change is made in air flow rate.
Because the air flow setpoint is derived from the fuel flow measurement, air flow will be affected by a false fuel flow measurement. During the load ramp considered above, if the fuel flow measurement over-reads by 20 %, the same quantity of additional air will enter the boiler. Since the fuel flow measurement is incorrect, there is no fuel to consume the oxygen in the additional air. Consequently, the 0 2 measurement will increase, and the 0 2 controller will start responding by reducing its output. The air flow setpoint will be reduced continuously by the 0 2 controller until the additional air flow has been eliminated and the 02 measurement is on setpoint. 182
02 set point 0 2 controller 02 measurement [4(
Furnace Air flow measurement
Air flow controller
Air flow set point Forced daught fan
Fuel flow measurement
Figure 7.25 Air flow and 02 control.
At the end of the load ramp, the awl flow will stabilize, the mill dynamics will expire and the fuel flow measurement signal will reduce to the true value of fuel flow. The air flow will be reduced in proportion with the fuel flow measurement. All this happens while the real fuel flow remains virtually constant. The reduction in air flow with constant fuel flow then reduces the 02 concentration in the flue gas to the normal value.
This happens under normal boiler control and it also happened under OHD control. Under these conditions, the OHD controller could not effectively manipulate the 0 2 - the influence from the incorrect fuel measurement was too strong. An attempt was made to speed up the 02 controller, but the limit of stability was reached before any improvement was noticeable.
A second method was devised which took into account the inability of the OHD to influence the furnace air flow by manipulating the 0 2 setpoint. Two optimization runs were done with this method. The first run was made to obtain the desired 0 2 setpoint. The second run was made with the 0 2 input to the model fixed to the actual measured 0 2 concentration in flue gas. The optimizer then ran and optimized the heat transfer rate the 183
best it could without changing the 0 2 setpoint. The final control values that were output
by the OHD control system to the P13 system were the 0 2 setpoint obtained from the first
optimization run and the other control element setpoints obtained from the second run.
Although this method showed some improvement in heat transfer rate to the reheater when
it was tested on the spreadsheet heat transfer model, in practice it introduced large process
oscillations. The difference between the model and the real plant (in this perspective) is
that on the actual plant, the 01-ID controller balances the excess heat transfer with injecting
additional spray water. During a load ramp, the 0, measurement increases due to the
reasons given above. The heat transfer predictor translates the increased 0 2 to excess heat transfer to the reheater.
During the first optimization run, the OHD optimizer takes action against the predicted
heat excess by reducing the setpoint to the 0 2 controller. During the second optimization
run, the 02 is not optimized, and a large degree of excess heat transfer to the reheater is
predicted. Consequently, the spray flow to the reheater is increased substantially. This
spray water is evaporated in the reheater and produces additional steam flow to the IP and
LP turbines. This increases the generator load output. The generator load controller
closes down the governor valves, thereby reducing the main steam flow and increasing the
boiler pressure. The pressure controller, in turn, reduces the boiler firing rate. This
reduces the error on fuel measurement, which reduces the furnace air flow. Consequently, the heat transfer to the reheater is reduced. This is reflected in the 0 2 concentration, and the OHD controller reduces the reheater spray flow rate - which starts the'same sequence in the opposite direction.
Figure 7.26 shows cycling this effect as recorded during a load ramp test from 586 MW to 486 MW at 15 MW/min with A, B, D, and E mills in service. The 0 2 deviations which result from the excess air (due to the untrue fuel flow measurement) are clearly evident.
The excess air increases the convective heat transfer rate. This effect is correctly predicted by the neural network model as excess heat discharged to the (mainly convective) reheater.
The deviations in heat transferred to the reheater are shown in Figure 7.27.
184
80 6
Ci(
60 2
Time (30 mintes) — Fuel flow rate 02 Concentration Steam flow rate
Figure 7.26 Deyiations in 02 measurement caused by incorrect fuel flow measurement.
240 .
7220
,t200
CO L3 180 1B' ±t ) 160
140
Time (30 mintes) Discharged — Target — Absorbed.
Figure 7.27 Effect on 02 on predicted heat discharge.
Deviations in heat transfer are balanced by the OHD controller through injection of reheater spray water. The resulting fluctuations in spay water flow rate are shown in Figure 7.28. Due to the undesired effect on process stability, the method of double optimization was removed from the OHD controller. Unfortunately, due to the poor control over furnace air flow rate, 0 2 setpoint manipulation was not a feasible means of controlling heat transfer with the current erroneous fuel flow measurement.
185
45 6
7)30) .z 4 0 0 a15 2
0 0
Time (30 minutes) — Reheat spray flow 02 Concentration
Figure 7:28 Reheat spray flow rate used by OHD control to absorb the excess heat transfer.
7.5 Final results OHD control was designed to reduce steam temperature excursions caused by load ramps and mill trips. The control philosophy was to predict the effect of fire-side disturbances on the process and then to calculate appropriate counter-acting control actions. Below are discussions on some OHD control aspects and on the results from some of the performance tests. Each test comprised
establishing reference test data with the normal unit controls , and then establishing performance data with OHD control active.
7.5.1 Bias action The OHD biassing action on the mills and burner tilts worked very well, apart from the oscillations caused due to the increased process gain that were sometimes evident. The 0, setpoint bias adjustment also worked well, but the air flow never really responded to this setpoint due to the fuel flow measurement errors. Figures 7.29 and 7.30 show the mill and burner tilt biassing recorded during a 150 MW load ramp from 536 MW to 686 MW at 15 MW per minute with A, B, C, & D mills in service. 186
100
90
80
70
60
50
40
Time (18 minutes) — A-Mill B-Mill C Mill — D Mill Normal
Figure 7.29 Biassed mill fuel flows under OHD control compared to normal.
Due to the excess heat entering the furnace during the upward ramp, the upper mills are biassed down in load, while the lower mills are biassed up to regulate heat flow to the superheater & reheater (Figure 7.29). Burner tilts are biassed downward to add to the heat shift (Figure 7.30) 0 2 biassing is not shown since OHD could not effectively manipulate it. During transients, 0 2 varied more with fuel flow than with setpoint changes.
30
20
10
a) 0
10
20
30
Time (18 minutes)
Normal — OHD
Figure 7.30 OHD tilt biassing during load ramp.
7.5.2 Controlling heat distribution
The biassing actions were generated to keep heat transfer rates to design. Improvements 187
in heat transfer rate were achieved on both superheater and reheater under 01-ID control. In most cases the regulation of heat transfer to the reheater was not as good as the superheater due to the lack of control over the furnace air flow rate. Heat transfer rates to the superheater and reheater recorded on a 150 MW downward load ramp from 686 MW to 536 MW at 15 MW/min with A, B, C, & D mills in service are shown in Figure 7.31 and Figure 7.32.
600
550
3 500
1; 450 C co 400
350
300 Time (30 minutes) — Target Normal — OHD
Figure 7.31 Heat transfer rate to superheater during down-ramp.
542
_540
538
g- 536 co E 534 co a) ° 532
530
Time (30 minutes) Normal — OHD
Figure 7.32 Effect of OHD control on main steam temperature. 188
7.5.3 Performance during load ramps
Load ramps are done on a daily basis to follow system load demands. The load ramp rate
is set at 15 MW/min. For this reason, load ramps during the evaluation of 01-1D control
were done at the same load ramp rate. Up and down ramps in load were done.
Up-ramps
During up-ramps in load, OHD control shifted the excess heat away from the superheater
and reheater to the evaporator. This assisted steam temperature control and deviations in
steam temperature were smaller with OHD control than without. Under conditions where
the OHD optimizer could not balance disturbances fully, the calculated increase in spray
water for balancing the remainder worked well. Cycling in process variables due to
increased gain in the pressure loop, were frequently evident. Results from a 200 MW load
ramp test from 486 MW to 686 MW at 15 MW/min with A, B, C, & D mills in service are
available in Appendix El.
Down-ramps
When active during down-ramps in load, OHD control shifted the excess heat away from the evaporator to the superheater and reheater. Unfortunately, manipulating the 0, setpoint proved largely unsuccessful due to the incorrect fuel flow measurement discussed earlier. With most tests, deviations in steam temperature were smaller with OHD control than without. Cycling was frequently evident. Results from a 100 MW load ramp test from 686 MW to 586 MW at 15 MW/min with A, B, C, & D mills in service are available in Appendix E2.
7.5.4 Performance during mill changes / trips
Better steam temperature control was not achieved with OHD control during mill trips and mill changes. The sudden, large shift in heat distribution resulted in large and quick movements of the control elements, and deviations in boiler pressure. Cycling between fuel and pressure then started and manual intervention was required. Results from -a test during which E-mill was tripped at 586 MW with A, C, & D mills remaining, are available in Appendix E3. 189
7.5.5 Performance during load runback During a capability load runback from full load, the unit load decreases almost instantaneously by 40%. One of the four mills in service is tripped automatically to assist with this sharp reduction in load. As with mill a normal trip, large changes in heat distribution occurs over a short period of time. Consequently, the process started cycling. No improvement in steam temperature control was achieved with OHD control during capability load runbacks. Unlike a mill trip, the fuel flow is fixed at 60 % after a unit capability runback. Steam pressure is then controlled by steam flow and not by fuel. Under OHD control, cycling still occurs, but with fuel flow fixed, the cycling occurs between steam flow and steam pressure.
A capability load runback test was done during by tripping one boiler water circulating pump to initiate the runbck. E-mill was tripped automatically and A, B, & D mills remained in service. Results from this test are available in Appendix E4. 190
8. Conclusion
8.1 Discussion
The thesis studied steam temperature control on power plant boilers. The role of power
generation in modern society was introduced and a historical overview of boiler controls was
given. It was reasoned that coal fired power stations will still be used for many years to come.
The mechanical and metallurgical importance of controlling steam temperature was motivated
(Chapter 1).
The power plant thermodynamic cycle was described, and three means of heat transfer between
fuel and boiler tubes were discussed: convection, radiation, and conduction. It was shown that
the balance between convective and radiant heat transfer changes through boiler load, while
conduction changes with boiler tube sooting. Reference was made to literature and it was
described how the placement and surface area of boiler components are critical to the design of
boilers. The sensitivity of heated elements to changes in heat distribution patterns was emphasized
(Chapter 2).
Various methods of steam temperature control and also the final control elements were described.
Three main classes of steam temperature control elements exist: altering the firing pattern,
changing the furnace air flow rate, and direct or indirect water cooling of steam. Long process
time lags, variations in process parameters, and process disturbances were identified as difficulties
associated with steam temperature regulation. Results from survey on steam temperature
excursions at Kendal were dicussed. Mill trips and load ramps, both causing fire-side disturbances,
were found to cause 80% of all excursions. The instrumentation and control configurations applied in practice were discussed and an overview of documented developments in advanced
steam temperature control on power plant boilers was made. Two main streams of progress were identified: model based / predictive control schemes and adaptive / nonlinear control schemes.
Comparative results between PID and advanced control showed definite benefits in applying advanced control methods to steam temperature control (Chapter 3).
The suitability of applying neural networks to process modelling and control were explored.
Neural networks were described and aspects related to the topology and training of networks were 191
discussed. It was argued that the nonlinear mapping capabilities and training properties of neural
networks are strong motivations for using neural networks to model existing processes. Various . neural network controller designs were described, and the error backpropagation technique was
shown to be well suited to the steam temperature control problem (Chapter 4).
The desired characteristics of a heat distribution model for a power plant boiler were listed. The
design and execution of a series of live plant tests for modelling data acquisition were explained.
Processing the data and calculating the heat transfer was described while all assumptions were
motivated. The calculation of many unmeasured variables was explained and specific attention was
given to discrepancies that appeared in the results. Using the 02 concentration in flue gas as an index of furnace air flow was motivated on the grounds of a very inaccurate air flow measurement.
The process of selecting the ideal network topology was described and comparative results were given. Improvements in modelling quality by selecting different model output schemes were shown. Modelling the heat transfer to boiler elements in relation to total heat discharge, with output adjustment to unity, was selected as the best modelling scheme on the grounds of results obtained (Chapter 5).
The requirements for improving steam temperature control were listed. It was showed that neural networks lend themselves very well to meet these requirements. The philosophy of optimal heat distribution (OHD) control was introduced. This scheme used plant measurements and a neural network heat transfer model to predict steam temperature excursions. The error backpropagation technique was then applied to the same neural network model to calculate the control actions necessary to prevent the excursions. In the case of optimizer or control element saturation, spray water quantities were calculated for eliminating the remaining errors (Chapter 6).
The 01-1D control algorithm was implemented on a personal computer and was interfaced to the boiler controls of an operational power plant. The development of the software programme was described and intricacies were pointed out. During the steady state testing phase, problems experienced with mill production rates and process noise were addressed. The optimization routine worked well and control elements were manipulated as expected. Transient tests showed an unexpected increase in process gain due to the control action manipulating the fireball inside 192 the furnace. This caused fuel-to-pressure oscillations which could not be eliminated effectively by decreasing the gain on the pressure controller. Erroneous fuel flow measurements during transient conditions affected the heat transfer calculations and air flow rate. Although the fuel flow signal could be improved for heat transfer calculations, the 0 2 setpoint could not be used effectively as a control element. Final results with OHD control were presented. Due to process oscillations caused by OHD control, a reduction in control quality was evident during mill trips and capability load runbacks. However, during load ramps, OHD control showed substantial improvements over normal PID control in main and reheat steam temperature regulation (Chapter 7).
8.2 Return to research hypothesis As part of the introductory Chapter, the hypotheses underlining the work done in this thesis, were stated. With these hypotheses in mind, the work done in this thesis may be concluded as follows:
The heat transfer from the firing system to the evaporator, superheater and reheater on a power plant boiler was effectively modelled by using a neural network trained on real plant test data. The best modelling results were obtained with a 7:5:3 neural network, modelling the heat transfer rate of individual components relative to the total heat transfer, and with error correction by adjusting outputs to summate to unity. Modelling accuracy was high and RMS errors were around 3.5 %.
This neural network model was used to estimate the effect that firing system disturbances would have on the boiler heat transfer before the steam temperature was affected significantly by these disturbances. Heat transfer rates were predicted and compared to design heat transfer rates. Any disturbances on the fire-side showed up instantaneously as errors on the comparators.
Adjustments to the firing system for minimizing the error between estimated heat discharge and design heat discharge were obtained from an optimization routine that iteratively backpropagated the errors through the neural network model. If the optimizer were unable to eliminate the errors entirely, corrective spray water calculations were done. 193
The new control scheme did not work well under disturbances caused by mill trips or load runbacks, due to process oscillations. However, during load ramps, the effect of firing system disturbances on steam temperature was reduced significantly.
To summarize the above points, the model predition obtained via a neural network was of high accuracy and could be used in a backpropagation control algorithm. However, stability aspects regarding the boiler pressure controller needs to be adressed.
8.3 Future research Future research should be aimed at improving the overall quality of ORD control. The two main areas needing attention are the accuracy of fuel flow measurement and stability of the pressure control loop.
8.3.1 Fuel flow measurement The accuracy of the fuel flow measurement is very poor during transients. The mill fuel flow feedback signal is not really a measurement but rather an estimation based on primary air flow rate and bypass damper position. A long term correction on the bypass damper is made when the indicated fuel flow rate and volumetric feeder speeds are mismatched. The mill feeders are driven by the mill level controller. Once again, the mill level is not measured but rather estimated from mill motor power level and sonic emissions from the mill drum. Both these measurements are also affected by the ball charge inside the mill. Complex, nonlinear, dynamic relations exist between the variables involved, and process parameters change through mill load and time.
The problem of estimating mill fuel flow and mill level could be possibly be solved to a large degree with a neural network model. Even if such a model is only about 90 % accurate, it will already reduce fuel flow indication errors by a factor of three. Apart from the improved fuel flow measurement, consequential advantages could be: improved air-fuel ratios during load ramps, improved pressure control during transients, and better furnace flame stability. 194
8.3.2 Stability of the pressure control loop Instability in the pressure control loop originated from the process gain increase due to the excess or deficit in heat transfer being directed to the evaporator. The gain increase is not constant and therefore the instability problem cannot easily be corrected just by recalculating the pressure controller settings.
Directing excess heat to the evaporator makes sense from a temperature control perspective, but it negatively influences the pressure control loop. This conflict in interests could be addressed by modelling the boiler pressure and temperature dynamic response. A recurrent neural network could be employed as the dynamic boiler model. Pressure and temperature targets can then be optimised simultaneously with a time-based minimum square error cost function. Backpropagation through time seems an ideal control solution for this expanded control scheme. In this case, a second recursive neural network may be used as a controller, but it must be able to manipulate the total fuel flow, as well as the other control elements used for heat distribution control.
By assigning the task of total fuel flow control to a neural network controller, maintaining stability in the pressure loop becomes a function of this controller. Process gain changes as a result of the mill biassing and burner tilting will still occur, but the same controller responsible for these control actions is also tasked with maintaining stability. Stability must therefore be one of the criteria built into the cost function to be minimised by the backpropagation algorithm. Consequently, the ability to maintain dynamic stability will be trained into the neural network controller. Success in this field has already been demonstrated by Nguyen and Widrow [115].
In this way, a new dynamic O1-ID controller will manipulate the amount of heat discharged, as well as its distribution, to control boiler pressure and steam temperatures simultaneously and with greatly improved stability. 195
Bibliography
Microsoft Corporation, "Electrical Power Systems," Microsoft Encarta 96 Encyclopedia, 1995. Eskom, Eskom annual report - 1996, Eskom, 1997. Eskom, POLL/ report, 24 July 1996. J. G. Singer, Combustion - Fossil Power, Combustion Engineering Inc, 1991. Siemens KWU, Siemens Offices Economic Data on Power Engineering - General Information, p. 35, Siemens Power Generation, 1993. 0. Mayr, The Origins of Feedback Control, MIT Press, 1970. S. Bennet, "A brief history of automatic control," IEEE Control Systems, pp. 17-25, June 1996. S. G. Dukelow, The Control of Boilers, Second Edition, Instrument Society of America, 1991. Central Electricity Generating Board, Modern Power Station Practice, Second Edition, Volume 6: Instrumentation, Controls and Testing, Ch.1, p. 1, Pergamon Press, 1971 Honeywell, "Introduction to distributed control systems," DCS/94 Seminar: Integrated process control, Laboratory for Advanced Engineering (PTY) LTD, 1994. J. Immelman, "Smart technology in pressure transmitters," SA Instrumentation & Control, pp. 34-35, July 1996. M. Barker, "Communication Busses," SA Instrumentation & Control, p. 38, April 1996. J. M. Birnbaum and J. P. Salkas, "DCS vs PLC: the conflict is over", SA Instrumentation & Control, pp. 6-13, February 1995. M. Barker, Special feature on Controllers, SA Instrumentation & Control, pp. 58-63 , July 1995. International Electrotechnical Commission, Specification for Steam Turbines, Publication 45, WC, Geneve, 1970. ABB, "Pressure Part Life Analysis," Report on the Kendal Boilers' Tubing, Header and Piping life assessment, ABB-Combustion Engineering, September 1993. S. H. Avner, Introduction to physical metallurgy, Second edition, McGraw-Hill, 1974. ASME, The ASME Boiler and pressure vessel code, Section 1, Power Boilers, 1989. G. H. Lausterer and G. Kallina, "Advanced unit control within stress limits," ASME Joint 196
International Power Generation Conference, Phoenix, AZ, October 1994. H. C. Le, "Start up and shutdown requirements for Escom Kendall #1 as controlled by thick walled component," Engineering Development Department, Combustion Engineering Inc., March, 1987.
G. J. Van Wylen and R. E. Sonntag, Funamentals of Classical Thermodynamics, John Wiley & Sons, 1985.
S. N. L. Carnot, E. Clapeyron and R. Clausius, Reflections on the motive power of fire; and other papers on the second law of thermodynamics. Peter Smith, 1962. W. J. M. Rankine, A manual of the steam engine and other prime movers, revised by W.J. Millar, Griffin and Co., 1908.
J. K. Salisbury and A: Calif, "Analysis of the steam-turbine reheat cycle," Transactions of the ASME, Vol 80, pp. 1629 - 1642, November 1958. M. G. Guile, "Light model method for simulation of the radiant heat flux distribution in the combustion chamber," Flow, processing and heating techniques, pp. 43 - 48, University of Novi Sad, Yugoslavia, 1975.
J. P. Holman, Heat Transfer, SI Metric Edition, McGraw-Hill, 1989.
J. P. Holman, Thermodynamics, 3rd ed., p. 350, McGraw-Hill, 1980. Babcock & Wilcox, Steam: its generation and use, 39th Edition, The Babcock and Wilcox Company, 1978.
Central Electricity Generating Board, Modern Power Station Practice, Volume 2:
Mechanical (boilers; fuel - and ash -handling plant), Pergamon Press, 1971.
F. P. Incropera and D. P. DeWitt, Introduction to heat transfer, 3rd Edition, John Wiley & Sons, 1996.
G. F. Hewitt, G. L. Shires and T. R. Bott, Process heat transfer, CRC Press Inc., 1994. M. Breedske, Discussions held after his six month training period at ABB Combustion Engineering design offices in Windsor USA, Kendal Power Station, June 1996.
C. D. Shields, Boilers: Types, Characteristics, and Functions, McGraw - Hill, 1961.
B. G. Liptak, "Optimization of Boilers," Optimization of Unit Operations, Chilton Book Company, 1987. J. A. Makuch, "Superheat steam temperature excursions," Report to Kendal Project Manager, Combustion Engineering, September 1991. 197
N. L. Baruffa, Load Cycling of the Kendal Boilers, Technology leadership programme dissertation, Eskom, 1992. Central Electricity Generating Board, Modern Power Station Practice, Third Edition, Volume 13: Boilers and Ancillary Plant, Ch.7, Pergamon Press, 1991 The International Society for Measurement and Control, • Fossil Fuel Plant Steam Temperature Control System - Drum Type, American National Standard ANSI / ISA - S77.44-1995, ISA, 1995. L. Aitchison and D. Bruggeman, "Ontario Hydro Nanticoke Generating Station ACORD Advanced Steam Temperature Control Test Results and Implementation," Instrumentation in the Power Industry, ISA Proceedings, Vol 35, pp. 177-195, 1992. H. Nakamura and M. Uchida, "Optimal Regualtion for Thermal Power Plants," IEEE Control Systems Magazine, Vol. 9, No. 1, pp. 33-38, 1989. D. W. St. Clair, Controller Tuning and Control Loop Performance, Second Edition, Straight-Line Control Company, Incorporated, 1995. F. G. Shinskey, Process Control Systems: Application, Design, and Tuning, McGraw-Hill, 1988. S. R. Valsalam, "Optimal Prediction and Control of Steam Temperature in Thermal Power Plants," Modelling, Simulation & Control, B, Vol. 25, No. 4, pp. 1-13, ASME Press, 1989. ABB Control and Monitoring System for Power Stations, Procontrol P, Kendal Power Station Control Diagrams, Issue 3, 1994. S. Zhu, M. Pantele and D. Labbe, "Chesterfield 6: An Evaluation for Model-Based Control," Instrumentation in the Power Industry, Proceedings, Vol.35, pp. 235-252, 1992. Kendal Boiler Design Manual, Combustion Engineering (now ABB), 1987. C. J. Franchot, "Implementation of a Model Based (MRFF) control Strategy for Steam Temperature Control of Cycling Power Plants," Proceedings of the American Power Conference, Vol. 56, No. 1, pp. 831-836, 1994. A. B. Corripio, Tuning of Industrial Control Systems, Instrument Society of America, 1990. S. A. W. M. Peters and W. W. de Boer, "Optimal Control for the Optimization of a Conventional Steam Temperature Controller," IFAC Symposium Series, No. 9, pp. 109- 198
114, 1992. R. Swanekamp, "El Segundo pioneers use of multivariable control technology," McGraw- Hill 's Power Magazine, pp. 72-76, June 1994. J. Anderson, "Control Strategies for Processing," C&I Europe, p. 33, October 1996. A. R. March, "Oldbury Power Station Steam Terhperature Control System", IEE Colloquium (Digest), No. 6, pp. 3/1 - 3/6, 1984. J. R. Riggs, K. Curtner, A. Agarval, "Comparison of two Advanced Steam Temperature Control Algorithms for Coal Fired Boilers," Instrumentation in the Power Industry, pp. 559-567, 1991. K. L. Hitz, P: Johnman and R. A. Cockerell, "Upgrade of Steam Temperature Control at Eraring Power Station, NSW," Instrumentation in the Power Industry, Proceedings, Vol: 35, pp. 197-205, 1992. S. Matsumura, K. Ogata, S. Fujii, H. Shioya, and H. Nakamura, "Adaptive Control for the Steam Temperature of Thermal Power Plants", Control Engineering Practice, Vol. 2, No. 4, pp. 567-575, 1994. B. W. Surgenor and J. K. Pieper, "An Optimal Multivariable Controller with application to Steam Temperature Control in a Boiler,", Transactions of the ASME - Journal of Dynamic Systems, Measurement, and Control, Vol. 114, pp. 733-736, 1992. K. Hanson, J. Werre, J. Chloupek, and J. Richerson, "Model-based control rescues boiler from steam temperature excursions," McGraw-Hill's Power Magazine, Vol. 139, No. 5, pp. 66-68, May 1995. N.K. Sinha, Control Systems, CBS College Publishing, 1986. R. Hertzog and F. Laubli, "Observer Control for Superheater Arrangements with Less Injection Points", Sulzer Akteiengesellschaft, Winterthur, Switzerland, 1987. K. Kt-tiger, "Process-Engineering Optimisation and Efficiency Increase by Advanced Control Methods," File: mAdaten\lcni\mw\cne96.doc. ABB Kraftwerksleittechnik GMBH, Manheim, 1996. S. Sastry, Adaptive Control: Stability, Convergence, and Robustness, Prentice-Hall International, Inc., 1989. W. D. T. Davies, System Identification for Self-Adaptive Control, John Wiley & Sons Ltd, 1970. 199
J. F. Smuts, "A qualitative comparison between PID, adaptive and neural network control
with reference to applications in drum level control on nonlinear plant," M.Eng. Dissertation, Rand Afrikaans University, January 1994.
M. Nomura and Y. Sato, "Adaptive Optimal Control of Steam Temperatures for Thermal
Power Plants," IEEE Transactions on Energy Conversion, Vol 4, No. 1, March 1989.
S. Haykin, Adaptive Filter Theory, Second Edition, Prentice-Hall, 1991. T. Xiangchu and T. Chengyuan, "A New Approach to Fuzzy Control," Fuzzy Logic in Knowledge-Based Systems , Decision and Control, by M. M. Gupta and T. Yamakawa,
pp. 307-315, Elsevier Science Publishers B.V., 1988.
Siemens, "Fuzzy Logic Controllers," Teleperm XP, Siemens Power Generation, 1996. J. A Bernard, "Use of a Rule-Based System for Process Control," IEEE Control Systems Magazine, Vol. 8, No. 5, 1988.
X. Peng, S. Liu, T. Yamakawa, P. Wang, and X. Liu, "Self Regulating PID Controllers and its Applications to a Temperature Controlling Process," Fuzzy Computing, M.M. Gupta, T.. Yamakawa, Elsevier Science Publishers, pp. 355-364, 1988. J. A. Rovnak, "Dynamic Matrix Based Control of Fossil Power Plants," IEEE Transactions on Energy Conversion, Vol. 6, No. 2, pp. 320-326, 1991.
M. P. Visser, "Turbine Throttling on Down Ramps," Minutes of meeting, Kendal Projects
Department, 23 March 1993.
R. Lippman, "An introduction to computing with neural networks", IEEE ASSP Magazine, Vol. 4, pp. 4-22, 1987.
W.S. McCulloch and W. Pitts, "A logical calculus of ideas immanent in nervous activity", Bulletin of Mathematical Biophysics, Vol. 5, pp. 115-133, 1943.
S.P. Chitra, "Use Neural Networks for Problem Solving", Chemical Engineering Progress, April 1993.
V.R. Vemuri, Artificial neural networks: Concepts and control applications, IEEE
Computer Society Press, 1992.
W.T. Miller, R.S. Sutton, and P.J. Werbos, Neural Networks for Control, MIT Press,
1990.
B.G. Liptak, Instrument Engineer's Handbook Volume 2: Process Control, Third Edition,
The Chilton Book Company, 1994. 200
K. Hornik, M. Stincombe, and H. White, "Universal Approximation of an Unknown Mapping and its Derivatives using Multilayer Feedforward Networks," Neural Networks Vol 3, pp. 211-223, 1990. A. Draeger, S. Engell, and H. Ranke, "Model Predictive Control Using Neural Networks," IEEE Control Systems Magazine, pp. 61-66, October 1995. D.L. Bailey and D.M. Thompson, "How to develop neural network applications," AI Expert, pp. 38-47, June 1990. D. Hillman, "Software review: Neuroshell v3.2," AI Expert, pp. 61-64, September 1990. Neural Ware Inc., Neural Computing, Technical Publications Group, Pennsylvania, 1991. Brainmaker Professional, California Scientific Software, 1992. K.S. Narendra and K. Parthasarthy, "Identification and Control of Dynamical Systems Using Neural Networks," IEEE Transactions on Neural Networks, Vol. 1, No. 1, pp. 4-26, March 1990. (85] K.L. Moore, Iterative learning for deterministic systems, Springer-Verlag London Limited, 1993. K.F. Reinschmidt, "Neural networks: next step for simulation and control", Power Engineering International, February 1993. K.B. Lee, "Practical aspects of multivariable statistical modelling of a boiler system for an electric power generating unit," IFAC Proceedings Series, No. 9, pp. 181-187, 1989. G. Klefenz and R. Seidel, "Control Characteristics of a Power Plant Unit operated under Controlled Sliding Pressure," VGB Kraftwerkstechnik, 56th Annual series, No. 2, pp. 75- 82, 1976. G. Pellegrinetti and J. Bentsman, "Nonlinear Control Oriented Boiler Modelling - A Benchmark Problem for Controller Design", IEEE Transactions on Control Systems Technology, Vol. 4, No. 1, pp. 57-64, January 1996. P.K. Chawdry, B.W. Hogg, "Identification of boiler models," IEE Proceedings, Vol. 136, Pt. D, No. 5, pp. 261 - 271, September .1989. L. Wang, "Design and Analysis of Fuzzy Identifiers of Nonlinear Dynamic Systems," IEEE Transactions on Automatic Control, Vol. 40, No. 1, January 1995. B. Fernandez, A.G. Parlos, and W.K. Tsai, "Nonlinear Dynamic System Identification using Artificial Neural Networks," International Joint Conference on Neural Networks, 20l
No. 90, Section II, pp. 133-141, 1990. K.S. Narendra, "Adaptive Control Using Neural Networks," Neural Networks for Control, edited by W.T. Miller, R.S. Sutton, and P.J. Werbos, pp. 115-142, MIT Press, 1990. G. Irwin, M. Brown, B. Hogg, and E. Swidenbank, "Neural network modelling of a 200 MW boiler system," IEE Proceedings on Control Theory Applications, Vol. 142, No. 6, pp. 529-536, November 1995. V.V. Murty, R. Sreedhar, B. Fernandez, and G.Y. Masada, "Boiler System Identification using Sparse Neural Networks," Advances in Robust and Nonlinear Control Systems, DSC-Vol. -53, pp. 103-111, ASME, 1993. M.J. Willis, C. Di Massimo, G.A. Montague, M.T. Tham, and A.J. Morris, "Artificial neural networks in process engineering," IE E Proceedings, Vol. 138, Pt. D, No. 3, pp.256- 266, May 1991. G. Martin, "Soft Sensors get the Process Data you really want," Control and Instrumentation, p. 45, January 1997. M. Rao, Q. Xia, and Y. Ying, "Modelling via artificial neural network," Modelling and Advanced Control for Process Industries, pp. 245-263, Springer-Verlag, 1994. [99.] Pegasus Technologies Ltd., "Neural Network Solutions for Clean, Efficient Operations," Power, p. 56, January / February 1997. Fisher-Rosemount, "Measure the key process variables you thought were unmeasurable," Control & Instrumentation, p. 40, May 1996. Pegasus Technologies Ltd., "Penn Power Unit drops NO x while improving heat rate," Power Engineering, pp. 52-54, March 1997. P.J. Werbos, "An overview of Neural Networks for Control", IEEE Control Systems magazine, pp. 40-41, January 1991. M. Khalid, R. Yu&A and S. Omatu, "Neural Networks for Self-Learning Process Control Systms," Journal of Artificial Neural Networks, Vol. 1, No. 1, pp. 23-49, 1994. [ I04] J.L. Dirion, B. Ettedgui, M. Cabassud, M.V. Le Lann, and G. Casamatta, "Elaboration of a neural network system for semi-batch reactor temperature control: an experimental study," Chemical Engineering and Processing, No. 35, pp. 225-234, 1996. [105] K.S. Narendra and S. Mukhopadhyay, "Adaptive Control of Nonlinear Multivariable Systems using Neural Networks," Neural Networks, Vol. 7, No. 5, pp. 737-752, 1994. 202
V. Etxebarria, "Adaptive control of discrete systems using neural networks," IEE
Proceedings on Control Theory Applications, Vol. 141, No. 4, pp 209 - 215, July 1994. L. Jin, P.N. Nikiforuk, and M.M. Gupta, "Adaptive control of discrete-time nonlinear systems using recurrent neural networks," IEE Proceedings on Control Theory Applications, Vol. 141, No. 3, pp 169-176, May 1994. M. Khalid, S. Omatu, and R. Yusof, "Temperature regulation with neural networks and alternative control schemes," IEEE Transactions on Neural Networks, Vol. 6, No. 3, pp. 572 - 582, May 1995. S.I. Mistry and S.S. Nair, "Identification and Control Experiments using Neural Designs,"
IEEE Control Systems Magazine, pp. 48 - 57, June 1994. A.G. Barto, R.S. Sutton, and C.W. Anderson, "Neuronlike adaptive elements that can solve difficult learning control problems," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 13, pp. 835-846, 1983. D. Sbarbaro-Hover, D. Neumerkel, and K. Hunt, "Neural Control of a Steel Rolling Mill,"
IEEE Control Systems Magazine, pp. 69 - 75, June 1993. H.W. Joubert, P.L. Theron, and T. Lange, "The use of Neural Networks for Gas Loop
Process Optimisation," SA Instrumentation & Control, pp. 44-51, September 1996. P.J. Werbos, "Maximizing Long-Term Gas Industry Profits in Two Minutes in Lotus Using Neural Network Methods," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 19, No. 2, pp. 315-333, March/April 1989. T.R. Niesler and J.J. du Plessis, "Time-Optimal Control by means of Neural Networks,"
IEEE Control Systems Magazine, pp. 23 - 33, October 1995. D.H. Nguyen and B. Widrow, "Neural networks for self-learning control systems," International Journal on Control, Vol. 54, No. 6, pp. 1439-1451, 1991. P.J. Werbos, "Backpropagation Through Time: What it Does and How to Do It," Proceedings of the IEEE, Vol. 78, No. 10, October 1990. S. Bittanti and L. Piroddi, "GMV technique for nonlinear control with neural networks,"
IEE Proceedings on Control Theory Applications, Vol. 141, No. 2, pp 57 - 69, March
1994. K.J. Hunt, D. Sharbaro, R. Zbikowski, and P.J. Gawthrop, "Neural networks for control edited by M.M. systems - a survey," Neuro-Control Systems: Theory and Applications, 203
Gupta and D.H. Rao, IEEE Press, 1994. J.R. Jang and C.T. Sun, "Neuro-Fuzzy Modelling and Control," scheduled to appear in Proceedings of the IEEE, March 1995. Corel Quattro Pro, Version 7.0, Corel Corporation Limited, 1996. Borland C++ 5.0, Borland International Inc., 1996. E. Schmidt and U. Grigull, Properties of Water and Stean in SI-Units, Springer-Verlag, 1989. Kraftwerk Union, "Kendal Power Station, Heat Flow Diagrams," 1984. J. Z. Woszczyk, "An evaluation of the multipoint 'Matrix' oxygen measuring system", Kendal Power Station, May 1993. Procontrol P: System Description, BBC Brown Boveri Ltd (now Asea Brown Boveri), Power Plants, Publication No. CH-KW 1102 87 E, 1987. Windows New Technology (NT), Microsoft Corporation, 1994. Windows 95, Microsoft Corporation, 1995. Microsoft PLUS!, Microsoft Corporation, 1995. Quinn-Curtis Charting tools, Rev 2.5, Quinn-Curtis Inc., 1996. W. J. Peet and T. K. P. Leung, "Dynamic Simulation Application in Modern Power Plant Control and Design," IEE Conference Publication, Vol: 1, No. 388, pp. 121 - 129, 1994. 204
Appendix A. Heat distribution test programme
14-Feb-96 Test 1 5-Mill Tests: ABCDE Wednesday Unit load Mill on hand Mill setp Tilt pos % 02 setp 08:00 - 09:00 Sub 1 662 E 104 52 6.0 09:00 - 10:00 Sub 2 619 B 53 3 5.7 10:00 - 11:00 Sub 3 612 C 85 67 3.0 11:00 - 12:00 Sub 4 585 D 86 70 3.6 12:00 - 13:00 Sub 5 574 A 45 99 2.6 13:00 - 14:00 Sub 6 570 C 70 32 5.8 14:00 - 15:00 Sub 7 543 C 64 8 5.7 15:00 - 16:00 Sub 8 538 A 48 8 5.1
15-Feb-96 Test 2 4-Mill Tests: ABCD Thursday Unit load Mill on hand Mill setp Tilt pos % OZ setp 08:00 - 09:00 Sub 1 685 D 106 49 4.9 09:00 - 10:00 Sub 2 644 A 57 21 5.1 10:00 - 11:00 Sub 3 604 D 104 95 3.1 11:00 - 12:00 Sub 4 572 B 63 99 4.7 12:00 - 13:00 Sub 5 560 B 61 51 5.3 13:00 - 14:00 Sub 6 473 B 66 72 3.5 14:00 - 15:00 Sub 7 449 C 71 74 4.0 15:00 - 16:00 Sub 8 432 D 73 68 3.8
16-Feb-96 Test 3 4-Mill Tests: ABCE Friday Unit load Mill on hand Mill setp Tilt pos % OZ setp 08:00 - 09:00 Sub 1 683 C 94 42 3.6 09:00 - 10:00 Sub 2 641 C 99 19 3.7 10:00 - 11:00 Sub 3 611 B 57 67 3.5 11:00 - 12:00 Sub 4 587 B 77 83 5.8 12:00 - 13:00 Sub 5 524 B 70 3 3.5 13:00 - 14:00 Sub 6 493 C 86 76 5.0 14:00 - 15:00 Sub 7 439 C 67 26 4.1 15:00 - 16:00 Sub 8 402 A 46 _ 49 4.4 205
17-Feb-96 Test 4 4-Mill Tests: ABDE Saturday Unit load Mill on hand Mill setp Tilt pos % 02 setp 08:00 - 09:00 Sub 1 684 E 108 14 5.2 09:00 - 10:00 Sub 2 573 D 83 17 4.9 10:00 - 11:00 Sub 3 542 B 64 39 3.3 11:00 - 12:00 Sub 4 528 B 64 96 3.7 12:00 - 13:00 Sub 5 522 D 90 51 4.8 13:00 - 14:00 Sub 6 485 E 103 70 2.6 14:00 - 15:00 Sub 7 472 A 50 55 3.3 15:00 - 16:00 Sub 8 406 D 56 50 4.6
18-Feb-96 Test 5 4-Mill Tests: ACDE Sunday Unit load Mill on hand Mill setp Tilt pos % 02 setp 08:00 - 09:00 Sub 1 666 D 95 69 4.1 09:00 - 10:00 Sub 2 632 C 59 28 3.1 10:00 - 11:00 Sub 3 632 C 76 10 4.1 11:00 - 12:00 Sub 4 614 A 51 12 5.6 12:00 - 13:00 Sub 5 546 D 91 95 5.6 13:00 - 14:00 Sub 6 505 • D 79 71 3.5 14:00 - 15:00 Sub 7 489 C 69 63 4.6 15:00 - 16:00 Sub 8 441 A 48 8 4.1
19-Feb-96 Test 6 4-Mill Tests: BCDE Monday Unit load Mill on hand Mill setp Tilt pos % 02 setp 08:00 - 09:00 Sub 1 689 C 84 39 4.0 09:00 - 10:00 Sub 2 675 C 81 9 2.7 10:00 - 11:00 Sub 3 631 E 103 46 2.5 11:00 - 12:00 Sub 4 614 C 60 90 3.8 12:00 - 13:00 Sub 5 611 C 66 60 3.9 13:00 - 14:00 Sub 6 536 C 74 77 3.6 14:00 - 15:00 Sub 7 498 C 57 84 4.8 15:00 - 16:00 Sub 8 436 D 66 56 5.2 206
20-Feb-96 Test 7 3-Mill Tests: ACD Tuesday Unit load Mill on hand Mill setp Tilt pos % 0, setp 08:00 - 09:00 Sub 1 646 D 109 58 5.2 09:00 - 10:00 Sub 2 611 A 79 29 3.8 10:00 - 11:00 Sub 3 552 A 58 79 5.4 11:00 - 12:00 Sub 4 526 D 102 63 5.9 12:00 - 13:00 Sub 5 514 C 84 21 3.1 13:00 - 14:00 Sub 6 482 D 109 18 4.5 14:00 - 15:00 Sub 7 458 D 109 9 2.8 15:00 - 16:00 Sub 8 _ 400 C 80 29 2.6
21-Feb-96 Test 8 3-Mill Tests: BCD Wednesday Unit load Mill on hand Mill setp Tilt pos % 0, setp 08:00 - 09:00 Sub 1 613 B 80 14 5.0 09:00 - 10:00 Sub 2 603 B 78 90 5.4 10:00 - 11:00 Sub 3 576 B 69 38 - 5.1 11:00 - 12:00 Sub 4 544 D 102 54 3.6 12:00 - 13:00 Sub 5 509 C 77 99 3.2 13:00 - 14:00 Sub 6 462 . C 77 42 3.9 14:00 - 15:00 Sub 7 378 C 71 66 3.1 15:00 - 16:00 Sub 8 356 _ D 81 89 3.5
22-Feb-96 Test 9 3-Mill Tests: BCE Thursday Unit load _Mill on hand Mill setp Tilt pos % 0, setp 08:00 - 09:00 Sub 1 531 _ C 64 56 4.4 09:00 - 10:00 Sub 2 482 B 47 69 5.9 10:00 - 11:00 Sub 3 427 E 105 20 4.3 11:00 - 12:00 Sub 4 339 E 72 85 3.5 12:00 - 13:00 Sub 5 323 C 53 4 2.5 13:00 - 14:00 Sub 6 310 C 52 49 3.1 14:00 - 15:00 Sub 7 300 C 49 49 4.8 15:00 - 16:00 Sub 8 299 E 50 88 3.6 207
23-Feb-96 Test 10 3-Mill Tests: BDE _ Friday Unit load Mill on hand Mill setp Tilt pos % 02 setp 08:00 - 09:00 Sub 1 647 B 102 13 5.5 09:00 - 10:00 Sub 2 621 D 103 47 4.6 10:00 - 11:00 Sub 3 615 E 106 82 5.0 11:00 - 12:00 Sub 4 504 D 74 16 5.9 12:00 - 13:00 Sub 5 440 D 77 56 3.1 13:00 - 14:00 Sub 6 420 D 78 43 4.6 14:00 - 15:00 Sub 7 401 6 55 81 5.3 15:00 - 16:00 Sub 8 310 B 46 _ 49 _ 4.8
24-Feb-96 Test 11 3-Mill Tests: CDE Saturday Unit load Mill on hand Mill setp Tilt pos % Oz setp 08:00 - 09:00 Sub 1 640 D 104 82 4.9 09:00 - 10:00 Sub 2 567 C 60 34 3.9 10:00 - 11:00 Sub 3 549 E 105 0 4.7 11:00 - 12:00 Sub 4 524 C 46 70 5.2 12:00 - 13:00 Sub 5 481 D 91 13 2.6 13:00 - 14:00 Sub 6 450 D 81 21 4.7 14:00 - 15:00 Sub 7 420 D 78 43 4.6 15:00 - 16:00 Sub 8 333 E 68 16 5.0
25-Feb-96 Test 12 2-Mill Tests: BC Sunday Unit load Mill on hand Mill setp Tilt pos % 02 setp 08:00 - 09:00 Sub 1 392 B 78 54 5.1 09:00 - 10:00 Sub 2 378 B 79 38 5.6 10:00 - 11:00 Sub 3 363 B 54 60 3.9 11:00 - 12:00 Sub 4 356 C 91 33 3.9 12:00 - 13:00 Sub 5 338 C 93 1 4.8 13:00 - 14:00 Sub 6 336 C 88 68 2.5 14:00 - 15:00 Sub 7 300 B 50 58 4.2 15:00 - 16:00 Sub 8 296 _ B 51 _ 83 _ 3.1 208
26-Feb-96 Test 13 2-Mill Tests: BD Monday Unit load Mill on hand Mill setp Tilt pos % 0, setp 08:00 - 09:00 Sub 1 424 B 101 19 • 4.2 09:00 - 10:00 Sub 2 419 D 106 79 5.6 10:00 - 11:00 Sub 3 415 B 92 12 2.8 11:00 - 12:00 Sub 4 397 D 104 57 3.6 12:00 - 13:00 Sub 5 396 D 104 10 5.5 13:00 - 14:00 Sub 6 393 D 94 1 5.5 14:00 - 15:00 Sub 7 340 D 105 62 3.8 15:00 - 16:00 Sub 8 316 D 76 _ 75 4.0
27-Feb-96 Test 14 2-Mill Tests: CD Tuesday Unit load Mill on hand Mill setp Tilt pos % 02 setp 08:00 - 09:00 Sub 1 419 D 106 42 6.0 09:00 - 10:00 Sub 2 419 C 98 30 5.3 10:00 - 11:00 Sub 3 399 D 107 68 5.1 11:00 - 12:00 Sub 4 343 C 65 25 5.2 12:00 - 13:00 Sub 5 331 C 51 15 5.0 13:00 - 14:00 Sub 6 299 D 67 46 4.0 14:00 - 15:00 Sub 7 294' C 51 11 4.3 15:00 - 16:00 Sub 8 294 C 50 100 5.8
28-Feb-96 Test 15 2-Mill Tests: CE Wednesday Unit load Mill on hand Mill setp Tilt pos % 0, setp 08:00 - 09:00 Sub 1 414 E 106 58 2.9 09:00 - 10:00 Sub 2 391 C 88 51 5.0 10:00 - 11:00 Sub 3 386 E 107 4 3.3 11:00 - 12:00 Sub 4 372 E 109 58 3.7 12:00 - 13:00 Sub 5 347 C 56 23 4.1 13:00 - 14:00 Sub 6 344 C 71 74 2.7 14:00 - 15:00 Sub 7 329 C 71 6 5.8 15:00 - 16:00 Sub 8 307 E 77 _ 90 5.6 209
29-Feb-96 Test 16 2-Mill Tests: DE Thursday Unit load Mill on hand Mill setp Tilt pos % 0, setp 08:00 - 09:00 Sub 1 371 E 96 68 4.1 09:00 - 10:00 Sub 2 345 D 63 31 3.8 10:00 - 11:00 Sub 3 336 E 77 1 3.3 11:00 - 12:00 Sub 4 333 E 92 92 4.5 12:00 - 13:00 Sub 5 323 0 67 35 5.5 13:00 - 14:00 Sub 6 310 E 70 89 5.4 14:00 - 15:00 Sub 7 298 D 45 33 4.4 15:00 - 16:00 Sub 8 288 E 68 _ 49 4.5 210
Appendix B. Variables recorded during heat distribution tests
Pt. Variable name AP no. Signal address Tag Name FEED WATER & SPRAY ENTHALPY 1 DST Press AP0649 LAA10CP001 XQ01 DST STM PR 2 DST Temp AP0183 LAA1OCT001 XQ01 DST WTR TMP 3 BFP outlet Press AP0680 LAB40CP001 X1001 BFP COMMON OUT FW PR 4 BFP outlet Temp AP0802 LAB4OCT001AXQ 01 BFP COM OUT FW TMP 5 Total feedwater flow AP1117 LABOOCF901 ZQ01 TOTAL FW FL FEED HEATERS & HP EXTRACTION 6 HP htr 6X steam Press Use cold reheat press AP0392 7 HP htr 6X steam Temp AP0827 MAA5OCT021 XQ01 HP TRB EXH TMP 8 HP htr 6X dist Press Use cold reheat press AP0392 9 HP htr 61 dist Temp AP0638 LCH61CT001AXQ01 HP HTR 61 CND OUT TMP 10 HP htr 5X fwtr Press AP0763 LAB50CP001 XQ01 FW CTRL VLV DIS PR 11 HP htr 51 fwtr Temp AP0803 LAB51CT002 XQ01 HP HTR 51 FW OUT TMP 12 HP htr 6X fwtr Press Use fwcv disch press AP0763 13 HP htr 6X fwtr Temp AP0807 LAB6OCT001 XQ01 HP HTRS OUT FW TMP ECONOMIZER 14 Eco outlet Press Use drum press AP0006 15 Eco outlet Temp LH AP0596 HAC21CT401 XQ01 LH ECON OUT TMP 16 Eco outlet Press Use drum press AP1111 17 Eco outlet Temp RH AP0597 HAC22CT401 XQ01 RH ECON OUT TMP EVAPORATOR 18 Drum Press AP0006 HAD6OCP001 ZQ01 DRM PRESS 19 Drum Press AP1111 HAN53CP001 XQ01 DRUM PRESS SUPERHEATER 20 Shtr atpr 1 LH in Temp AP0010 HANSI CT001 XQ01 LH SHTR ATPR 1 INL TMP 1 21 Shtr atpr 1 LH vv1 pos AP1254 LAE61AA001 XQ50 LH SHTR ATPR 1 VLV 1 POS 22 Shtr atpr 1 LH vv2 pos AP0676 LAE63AA001 XQ50 LH SHTR ATPR 1 VLV 2 POS 23 Shtr atpr 1 LH out Temp AP0011 HANSI CT002 X001 LH SHTR ATPR 1 OUT TMP 1 24 Shtr atpr 1 RH in Temp AP0013 HAH52CT001 XCIO1 RH SHTR ATPR 1 INL TMP 1 25 Shtr atpr 1 RH vv1 pos AP 1255 LAE62AA001 XQ50 RH SHTR ATPR 1 VLV 1 POS 26 Shtr atpr 1 RH vv2 pos AP0017 LAE64AA001 XQ50 RH SHTR ATPR 1 VLV 2 POS 27 Shtr atpr 1 RH out Temp AP0014 HAH52CT002 XQ01 RH SHTR ATPR 1 OUT TMP 1 28 Shtr atpr 2 LH in Temp AP0731 HANSI CT011 XQ01 LH SHTR ATPR 2 INL TMP 1 29 Shtr atpr 2 LH vv1 pos AP0741 LAE81AA001 XQ50 LH SHTR ATPR 2 VLV 1 POS 30 Shtr atpr 2 LH vv2 pos AP0743 LAE83AA001 XQ50 LH SHTR ATPR 2 VLV 2 POS 31 Shtr atpr 2 LH flow AP0747 LAE91CF001 ZQ01 LH SHTR ATPR 2 SPRWTR FL 32 Shtr atpr 2 LH out Temp AP0733 HAH81CT013 XQ01 LH SHTR ATPR 2 OUT TMP 1 33 Shtr atpr 2 RH in Temp AP0735 HAH82CT011 XQ01 RH SHTR ATPR 2 INL TMP 1 34 Shtr atpr 2 RH vv1 pos AP0742 LAE82AA001 XQ50 RH SHTR ATPR 2 VLV 1 POS 35 Shtr atpr 2 RH vv2 pos AP0744 LAE84AA001 XQ50 RH SHTR ATPR 2 VLV 2 POS 36 Shtr atpr 2 RH flow AP0746 LAE92CF001 ZQ01 RH SHTR ATPR 2 SPRWR FL 211
37 Shtr atpr 2 RH out Temp AP0737 HAH82CT013 XQ01 RH SHTR ATPR 2 OUT TMP 1 38 Total Shtr atpr flow AP1253 LAE50CF001 ZQ01 D/SHTR SPRWTR FL 39 Shtr outlet Press LH AP1106 LBA11CP901 XQ01 LH SHTR OUT PR 40 Shtr outlet Temp LH AP1107 LBA11CT904 XT03 LH SHTR OUT TMP 41 Shtr outlet Press RH AP1164 LBA12CP901 XQ01 RH SHTR OUT PR 42 Shtr outlet Temp RH AP1108 LBA12CT904 XT02 RH SHTR OUT TMP 43 Main steam Flow AP0600 HAH8OCF900 XQ01 STM FL REHEATER 44 Rhtr inlet Press AP0392 LBC12CP401 XQ01 RH CRHT (HP EXHAUST) 45 Rhtr inlet Temp LH AP0327 LBC11CT001 X001 LH RHTR ATPR INL TMP 46 Rhtr atpr LH vv1 pos AP1259 LAF53AA001 XQ50 LH RHTR ATPR VLV 1 POS 47 Rhtr atpr LH vv2 pos AP1260 LAF55AA001 XQ50 LH RHTR ATPR VLV 2 POS 48 Rhtr atpr out Temp LH AP1119 LBC11CT003 XQ01 LH RHTR ATPR OUT TMP 49 Rhtr inlet Temp RH AP1118 LBC12CT001 XQ01 RH RHTR ATPR INL TMP 50 Rhtr atpr RH vv1 pos AP0323 LAF54AA001 XQ50 RH RHTR ATPR VLV 1 POS 51 Rhtr atpr RH vv2 pos AP0324 LAF56AA001 XQ50 RH RHTR ATPR VLV 2 POS 52 Rhtr atpr out Temp RH AP1120 LBC12CT003 XQ01 RH RHTR ATPR OUT TMP 53 Total Rhtr atpr flow AP1176 LAF40CF001 ZQ01 RHTR ATPR SPRWTR FL 54 Rhtr outlet Press AP1121 LBB22CP011 XQ01 IP ESV1 HRS INL PR 55 Rhtr outlet Temp LH AP1104 LBB11CT001 XQ01 LH RHTR OUT TMP 56 Rhtr outlet Temp RH AP1105 LBB12CT001 XQ01 RH RHTR OUT TMP 57 Rhtr outlet Temp Setpnt AP1262 LBB12DT901 XT04 RH RHTR TMP SETPNT FUEL & FIRING 58 Fuel flow A mill AP0138 HFE5ODU500 XT01 MILL A TOT FUEL FL 59 A-mill feeder speed DE AP0383 HFB52DS001 XQ50 MILL A DE FDR SPD 60 A-mill feeder speed NDE AP0382 HFB51DS001 XQ50 MILL A NDE FDR SPD 61 Fuel flow B mill AP0154 HFE4ODU500 XT01 MILL B TOT FUEL FL 62 B-mill feeder speed DE AP1251 HFB42DS001 XQ50 MILL B DE FDR SPD 63 B-mill feeder speed NDE AP1252 HFB41DS001 XQ50 MILL B NDE FDR SPD 64 Fuel flow C mill AP0174 HFE30DU500 XT01 MILL C TOT FUEL FL 65 C-mill feeder speed DE AP0605 HFB32DS001 XQ50 MILL C DE FDR SPD 66 C-mill feeder speed NDEAP0604 HFB31DS001 XQ50 MILL C NDE FDR SPD 67 Fuel flow D mill AP0007 HFE2ODU500 XT01 MILL D TOT FUEL FL 68 D-mill feeder speed DE AP1045 HFB22DS001 XQ50 MILL D DE FDR SPD 69 D-mill feeder speed NDEAP1044 HFB21DS001 XQ50 MILL D NDE FDR SPD 70 Fuel flow E mill AP0266 HFE1ODU500 XT01 MILL E TOT FUEL FL 71 E-mill feeder speed DE AP0607 HFB12DS001 XQ50 MILL E DE FDR SPD 72 E-mill feeder speed NDE AP0606 HFB11DS001 XQ50 MILL E NDE FDR SPD 73 Fuel oil flow AP0716 HJF00CF901 ZQ01 OIL FLOW 74 Total fuel flow AP0360 HFEOODU500 XT11 TOT FUEL FL 75 Burner tilt position AP0682 HHAO10E001 ZTO1 BURNER TILT POS AIR FLOW 76 Primary air Flow AP0371 HLB00DF900 XT10 TOT PA FL 77 Secondary air Flow LH AP0365 HLB10CF901 XG)02 LH FD AIR FLOW 78 Secondary air Flow RH AP0366 HLB20CF901 XQ02 RH FD AIR FLOW 79 Total air Flow AP0363 HLBOOCF901 a)02 TOT AIR FL 80 02 content LH AP0381 HNA12CQ001 XQ01 LH 02 CONTENT 212
81 02 content RH AP0362 HNA22CQ001 XQ01 RH 02 CONTENT 82 Precip air inlet Temp LH AP0775 HNA14CT904 ZQ01 LH P/CIP GAS INL TMP AVR 83 Precip air inlet Temp RHAP0776 HNA24CT904 ZQ01 RH P/CIP GAS INL TMP AVR MISCELANEOUS 84 Target load AP0669 CJAOODU590 XJ07 UNIT TARGET LOAD 85 Generated MW AP1147 CJAOODU450 XU15 GEN MW 86 Unit load setpoint AP0670 CJAOODU570 XU01 UNIT LOAD SETPNT 87 Turbine demand AP0672 CJAOODU460 XU53 UDC TURB LOAD DEMAND 88 Boiler demand AP0671 CJA00DU500 XJ23 UDC BLR DEMAND 89 Shtr Press setpoint AP0361 CJAOODU460 XU51 SHTR OUT PR SETPNT 90 HP governor vv position AP0454 MAA12CG001 Xia01 HP GOV V1 POS 91 IP governor vv position AP0455 MAB12CG001 XQ01 IP GOV V1 POS 92 Ambient air Temp AP0580 PADOOCT001 XQ01 CENT C/TOWR INL AIR TMP 93 Condensor vacuum AP0816 MAG10CP005 XQ01 COND 1 VAC 94 Dust level AP0715 HME1OCQ001 XQ01 DUST LEVEL 95 Fuel factor AP0666 CJAOODU540 XT12 FUEL FACTOR 213
Appendix C. Spreadsheet model
Previously in this thesis, mention was made of a boiler heat transfer model that was used to
determine heat transfer rates. This model was created on a Corel Quattro Pro [120] spreadsheet
running on a personal computer. The spreadsheet has a neural network as its core which it uses
to calculate heat transfer rates from any given set of boiler conditions It can also be used for
optimization of control elements to achieve desired heat transfer rates.
An engineering interface is used to input the state of various furnace elements. The modelled heat
transfer rates are calculated from these inputs and displayed numerically and graphically. All the
neural network calculations are done in spreadsheet cells.
Four spreadsheet pages are used by the heat transfer model, each with a specific set of
calculations. The first page is configured as the engineering interface. From here changes can be
made to mill firing rates, burner tilt angle, and 0, setpoint. Fuel factor and coal calorific value can
also be adjusted. The second page does scaling of all the variables for use by the neural network.
The latter is configured over two pages, one for each layer of neurons in the network. The neural
network outputs are rescaled to relative heat transfer rates on the second page and adjusted to add up to unity. The first page displays the modelled heat transfer rate to the evaporator, superheater, and reheater.
Brainmaker [83] was used to train the neural network. The training data was obtained during a series of special heat transfer tests run on Kendal Unit 3. The neural network weights were uploaded from a network configuration file created by Brainmaker after training.
The model can also be used for optimization. In this mode, target heat discharge rates for the evaporator, superheater, and reheater are entered into allocated celles. Errors in heat transfer can be weighted individually. Thereafter the built in optimizer of Quattro is used to manipulate furnace elements to obtain target heat transfer rates to the superheater and reheater. The optimizer minimizes the sum of the weighted RMS errors between the target heat transfer rates and the model outputs. Limits may be placed on total fuel flow rate, individual mill fuel flow rates, 0, 214 concentration, and burner tilt angle. Optimization is performed within these limits.
The heat transfer model, which is quite easy to operate, is used as an engineering tool for determining heat transfer rates under various conditions. Figure C.1 displays the engineering interface of the model.
I 1 I I INPUTS =r cr ' r .4 111114:0111S IN Mia BOILER HEAT GAINS = .., .--,-, Niagatimow . ..i, ., al12101111M NI ai as % of total heat transfer — pa', -my- IIMUILSE a ' IS/2a MI MIES I 0 1;3111111111111Sal Mill SI R-heat (18.37%) iiiirEann=mai ■ msMEMOISMILM 91 11111 11•4111,M=1. fl:11 S-heat (33.12%) : • - - EMILILIIIIIIII II ..7-iiiiiiia tuna . =Mel" =Aar, "0 •in ou" -9 21•1111111611052:1121r MISIM 111-11 Evap (48.51%) — - - . -..- wialilL■am a= IN • On uncorrectedibUrhonftranster , Figure C.1 The neural network boiler heat transfer model running on a spreadsheet. 215
Appendix D. OHD graphic display
12
r- a co N 0
Yr
0
a
■ • - • • • - • • ■ • ■ ■ ■ 0 0 0 0 0 0 0 CI CO CO CO 10 CO qv- E2
tics is ter ac har C fer s Tran
O t Hea iler Bo
0 E 8 8 i 0 8 lyric atsueineeN 216
Appendix E. Selected test results
The following four pages contain prints of the key parameters measured, calculated, and recorded from Kendal Unit 3 during the OHD control tests. Each page is dedicated to two tests, done under exactly the same conditions, one with only the normal boiler controls, and the other with OHD control active. The pages contain prints of test data recorded under the following conditions:
Appendix E 1 200 MW load ramp from 486 MW to 686 MW at 15 MW/min with A, B, C, & D mills in service.
Appendix E2 100 MW load ramp from 686 MW to 586 MW at 15 MW/min with A, B, C, & D mills in service.
Appendix E3 E-mill trip at 586 MW with A, C, & D mills remaining in sevice.
Appendix E4 Capability load runback from full load to 60 %. E-mill was tripped automatically and A, B, & D mills remained in service. One boiler water circulating pump was tripped to initiate the runbck. 217
Appendix El
200 MW load ramp from 486 MW to 686 MW at 15 MW/min with A, B, C, & D mills in service.
Superheater temperatures 550
545
540
535
530
525 — LH Norm — RH NOWT — LH w OHD — RH w OHD
Spray water flow rates Repeater temperatures 70 570
60 560 50 k \ / 40 550
30 -1 540 \ 20 / 530 10
0 520 — Star Norm —Rids Nona — Shtr w OHD — Rhtr w OHD — LH Norm — RH NOM — U-I w OHD — RH w OHD
Tilts & 02 OHD Mill biassing 30 120
20 -..'-f\--n 5 100 re- 10 ao \\ i / A A7C 0 3 so \ -{ '"i1 ,-, z \J i \! -10 • 2 40
20 20
-30 tionny 0 Tia NWT — 02 Norm — Titt w OHD — 02 w OHD - — 13-Mill — C-Mill — D-Mill — EMN
OHD Disabled OHD Enabled 800 900
700 800 700 600 600 500 500 400 400 300 300 200 200 100 100 — Discharge — Target — Absorbed — Discharged — Target — Absorbed 218 Appendix E2 100 MW load ramp from 686 MW to 586 MW at 15 MW/min with A, B, C, & D mills in service.
Total fuel & main steam flow rates Superheater temperatures 105 545
100 540 95
so 535 es \ k \ 530 NNy 8o
75 525 Fuel Non — Steam norm — Fuel OHD — Steam OHD LH Norm — RH Norm — LH w OHD — RH w OHD
Attemperatlon water flow rates Reheater temperatures 35
30 545 25 _ 20 r\A■ ft.____N 540 15 535 rilikalree‘ 10 . wq141rA,1 530 5 N 0 525 S1 NamM — Rhtr Norm — Star w OHD — Rhtr w OHD LH Noon — RH Nonn — LH w OHD — RH w OHD
OHD Mill biassing 100
80 so
40
20
0 — A-Mill — B-Mill — — 0-Mill — E-Mill
OHD Enabled 800
700
630 t/
500
400
300 I I • 200 — Discharged — Target — Absorbed 219 Appendix E3 E-mill trip at 586 MW with A, C, & D mills remaining in sevice.
Total fuel & main steam flow rates 95
91)
85 8o
75
70
55 — Fuel Norm — Steam norm — Fuel OHD — Steam OHD
Spray water flow rates Repeater temperatures 70 550
60 545
50 540
40 A 535 30 innenAleaw 530 telleM1111M-. 20 Italra•Maiena. 525 10 520
0 515 — Shtr Norm — Rhtr Norm — Shtr w OHD — RMr w OHD — LH Nom% — RH NOIM — LH w OHD — RH w OHD
Tilts & 02 OHD Mill biassing 30 6 120 -% A . Ai • 40 /a. • 20 5 100 AWAILTICatallalln NAN..4% allS112 10 ,c,itN a' .\ A 4 8o maka I \ ‘,/ 0 yansines 3 W 11 1 INi ° 11 10 2 40
-20 1 20
30 MV/Itallna 0 0 — TiS Noah —02 Norm — Till w OHD — 02 w OHD — A-Mill — &MID — GMia — DMiIl — E-Mill
OHD Disabled OHD Enabled eco 800 700 700 • ww - • .'tag , 600 600 • V 500 SW
400 400
300 300
200 200
100 103 — Discharged — Target — Absorbed — Discharged — Target — Absorbed
220 Appendix E4 Capability load runback from full load to 60 %. E-mill was tripped automatically and A, B, & D mills remained in service. One boiler water circulating pump was tripped to initiate the runbck.
Total fuel & main steam flow rates Superheater temperatures 110 545
100 540 535 90 530 ao 525 70 520 60 515 50 510 — Fuel Norm — Steam norm — Fuel OHD — Steam OHD — LH Norm — RH Norm — 1.11w OHD — RH w OHD
Spray water flow rates Reheater temperatures 50 550
ao 540
30 530 __ALVA■1111111ra 20 v 520
10 510
0 500 — Shtr Norm — Rhtr Norm — Sltr w OHD — Rhtr w OHD — LH Noun — RH Norm — LH w OHD — RH w OHD
Tilts & 02 30 6
20 5
10 4
0 3
10 2
-2o
-30 0 — Tilt NOM —02 Norm — Tilt w OHO — 02 w OHD
OHD Enabled SOO 700 600 500 400 300 200
100 — Discharged — Target — Absorbed