Optimization of Combustion Process in Coal-Fired Power Plant with Utilization of Acoustic System for In-Furnace Temperature Meas

Applied Thermal Engineering 123 (2017) 711–720

Contents lists available at ScienceDirect

Applied Thermal Engineering

journal homepage: www.elsevier.com/locate/apthermeng

Research Paper Optimization of combustion process in coal-ﬁred power plant with utilization of acoustic system for in-furnace temperature measurement ⇑ Łukasz S´ladewski a, , Konrad Wojdan a, Konrad S´wirski a, Tomasz Janda b, Daniel Nabagło b, Jerzy Chachuła b a Warsaw University of Technology, Institute of Heat Engineering, Nowowiejska 21/25 St., 00-665 Warszawa, Poland b EDF Polska CUW Sp. z o.o., Research & Development Department, Ciepłownicza 1 St., 31-587 Kraków, Poland highlights

Novel approach in combustion modelling and optimization in coal-fired boiler. R&D and implementation project that was carried on on real boiler. Artificial immune optimization system integrated with temperature profile measurement. Proper control of the fireball shape resulted in 0.27% boiler efficiency increase. article info abstract

Article history: This paper presents methodology and results of a research project on software optimization of combus- Received 26 November 2016 tion process efficiency in coal-fired power plant. Accepted 14 May 2017 The general goal of this project was to increase boiler efficiency by proper control of the combustion Available online 19 May 2017 process using optimization software, integrated with Distributed Control System and in-furnace temperature profile measurement system. The research goal relays on new approach in combustion modelling Keywords: based on in-furnace temperature distribution and utilization of this model in on-line boiler control. it is Combustion optimization assumed that this approach allows for more precise control of the combustion process, what finally has a Combustion advanced control positive influence on boiler performance – the efficiency in specific. Power boilers Boiler efficiency The solution has been designed, installed and tested on existing, utility plant – 225 MW (650 t/h of Acoustic temperature measurement nominal steam generation). Final analysis has shown positive results – the boiler efficiency increased over 0.25%. Ó 2017 Published by Elsevier Ltd.

1. Introduction modernization of boiler equipment e.g. new measurement technologies, to on-line software optimization systems. The electricity sector worldwide faces still-increasing demand During last few years contactless temperature measurement for cost efficient power generation and stricter environmental technologies have become more popular in the industry. There regulations. These two factors motivate operators, especially in are two main technologies, which are used in coal fired boilers to coal-fired generating stations, to search new solutions for process measure temperature distribution – acoustic technology [2–4] optimization. Nowadays, coal still plays an important role in and laser technology [5,6].In[7] authors present advantages of electricity production – over 40% of global electricity production acoustic system in evaluation of combustion quality in pulverized comes from coal [1]. That is why optimization of combustion pro- coal-fired boiler. On the other hand, in [8] the meaning of laser cess in terms of boiler efficiency and emission of air pollutants is technology in combustion optimization projects is presented. Tem- the key in minimizing operational and maintenance costs. There perature distribution in horizontal cross-section of a boiler is a are many methods for such optimization, which range from great indicator of quality of combustion process. This indicator could be used in boiler control to optimize the combustion process. The research key of this work is to develop new modelling ⇑ Corresponding author. approach of the combustion process which is based on in-furnace ´ E-mail addresses: [email protected] (Ł. Sladewski), konrad.wojdan@ temperature distribution. This new model will be used by combus- itc.pw.edu.pl (K. Wojdan), [email protected] (K. S´wirski), tomasz.janda@ edf.pl (T. Janda), [email protected] (D. Nabagło), [email protected] tion optimization software for on-line control of the combustion. (J. Chachuła). http://dx.doi.org/10.1016/j.applthermaleng.2017.05.078 1359-4311/Ó 2017 Published by Elsevier Ltd.

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 712 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720

Nomenclature

ðpÞ ai autoregressive coefficients T gas temperature [K] A current process operating point ua current value of process inputs vector i th AC Advanced Control ukj process input at k-j time stamp ½ =ð Þ k B universal gas constant J mol K , ub average value of process inputs vector before control iðpÞ th bj process input coefficients change in k lymphocyte ½ = k C speed of sound m s ua average value of process inputs vector after control cðpÞ constants change in kth lymphocyte CFD Computational Fluid Dynamics wðpÞ weights of disturbance partition ~ CV process outputs (Controlled Variables) yk Current value for k-th monitored process output D number of disturbance signals y ^k Demand value for k-th o monitored process output DCS Distributed Control System ya current value of process outputs vector r th DV process disturbances (Disturbance Variables) yk r process output in NAMAX method r th th FD forced draft fan yki r process output at k-i time stamp k g condition function for lymphocyte fitting process yb average value of process outputs vector before control ID forced draft fan change in kth lymphocyte k K number of inputs ya average value of process outputs vector after control th Kk matrix of the automatically identified input-output change in k lymphocyte gains of the process z process disturbances l distance between particular transmitter and particular za current value of process disturbances vector receiver [m] ak Linear penalty coefficient for k-th manipulated variable b LOI Loss on Ignition, k Square penalty coefficient for k-th manipulated variable c M molecular weight [mol] k Linear penalty coefficient for k-th monitored process M time horizon of process inputs output ~ c d mk Current value for k-th manipulated variable k Square penalty coefficient for k-th monitored process ^c m k Demand for k-th manipulated variable output d MIMO Multi-Input-Multi-Output D!m optimal change vector of manipulated variables MPC Model Predictive Control DT vector of optimal changes for twelve current AGAM MV process controlled inputs (Manipulated Variables) temperatures N autoregressive time horizon Duk control change in process inputs vector in kth lympho- p disturbance partition cyte P number of disturbance partitions Dyk static process response on outputs vector in kth lympho- _ Q sum of heat losses cyte _ loss Q sum of useful heat outputs lðL ; AÞ lymphocyte selection function _ out k Q tot sum of heat losses and useful heat outputs gB boiler efficiency R specific gas constant [J=ðkg KÞ] j adiabatic coefficient [–] R ! number of process outputs s flight time [s] ~ ð Þ slm sd T AGAM Dispersion of current shape, defined by vector of mea- k Insensibility zone for linear penalty for k-th manipu- sured AGAM temperatures lated variable ^ sm sd ! Dispersion in reference shape s Insensibility zone for square penalty for k-th manipu- ~ k siðT AGAMÞ Intensity of current shape, defined by vector of mea- lated variable sured AGAM temperatures sly Insensibility zone for linear penalty for k-th monitored ^ k si ! Intensity in reference shape process output ~ ð Þ ssy sp T AGAM Left-right position of temperatures hotspot of current k Insensibility zone for square penalty for k-th monitored shape, defined by vector of measured AGAM tempera- process output ðÞ ð Þ ¼ 1 ð þj jÞ tures þ ‘‘positive” operator x þ 2 x x ^ sp Left-right position of temperatures hotspot in reference shape

Regarding software optimization systems for advanced com- network and genetic algorithm for combustion process control bustion control there is number of different solutions. The main are described in [16–19]. Effective combustion optimization could difference between those solutions relay on difference in approach be also performed by solutions, which combine artificial neural in process modelling and difference in algorithms used to search network models and quadratic programming [20], artificial bee for optimal solution. MPC is the main group of advanced process colony [21] or particle swarm [22] algorithms. Moreover, in [22] control algorithms for MIMO processes e.g. combustion process. authors provide a comparative analysis of artificial neural network Detailed description of different MPC solutions are provided in and support vector machine models of the combustion process. [9,10] and their advantages in combustion optimization were The support vector machine algorithm [23] have found also its described in [11–14]. MPC as well as steady-state optimization application in combustion process optimization with genetic algo- solutions utilize often algorithms, which are inspired by processes rithm [24,25], ant colony [26,27], particle swarm, strength pareto that could be observed in the nature. One of most popular solutions evolutionary or archive-based hybrid scatter search [28] algo- combines artificial neural network algorithms for process mod- rithms. Computational fluid dynamics is another method for com- elling and genetic algorithm as the method for searching for the bustion process modelling, which has been used for on-line optimal solution. General idea of this solution could be found in combustion optimization. In [29,30] authors presented positive [15]. Additionally, examples of implementation of artificial neural effect of implementation of computational fluid dynamics model

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720 713 with genetic algorithm and particle swarm optimization methods. – Oxygen content in flue gasses, All of the presented algorithms succeeded in combustion optimiza- – Flue gas temperatures, tion projects – improved boiler efficiency, reduction of NOx emis- – Loss of ignition, sion, but also improved controllability of the process. All of the d Disturbances: solutions have also meaningful disadvantages, which are described – Boiler load, in chapter 4. – Configuration of operating mills, Other type of nature inspired optimization solutions are algo- – Fuel quality. rithms inspired by immune system. In [31] authors described architecture and combustion optimization results of SILO solution Combustion is a non-linear process. One of well-known mod- – the optimization algorithm, inspired by operation of immune elling methods of non-linear processes is NARMAX (non-linear system. autoregressive moving average with exogenous input) [37]. The In many projects, combustion optimization solutions rely on NORMAX relation between inputs and outputs of the non-linear standard DCS measurements. On the other hand, in [32–35] process presents the following equation. All the parameters are authors proved a great meaning of in-furnace temperature distri- defined basing on parametric tests that are performed on running bution in combustion process control. The [32,33] publication unit. describes acoustic technology and the [34,35] publication – laser ! XP X XK X technology. In publications, authors presented positive influence r ¼ ðpÞ r þ iðpÞ i þ ðpÞ yk wðpÞ ai yki bj ukj c ¼ 2 ¼ 2 of balanced temperature distribution on combustion process p 1 i N i 1 j Mi parameters, such as: excess air, CO and NOX emission, steam temperatures but especially – the boiler efficiency. Other examples, To increase the accuracy of the NORMAX model a neural net- where in-furnace temperature distribution and flame quality are work is introduced to minimize the error between model and pro- subjects of combustion optimization are described in [29,36,35]. cess outputs. In [29] the temperature distribution in the furnace is obtained from Artificial immune system could be considered as an alternative CFD simulation. The flame quality in [36] is quantified by analysis to the NORMAX method for combustion modelling. The artificial of high resolution flame images. immune system is the principle that was used in SILO [39] – the Advantages of SILO system and acoustic temperature measure- system implemented in this research project. In this method there ment technology – AGAM, motivated authors of this paper to carry are the same input, output and disturbance signals represented by on a research on novel method in combustion process modelling vectors. and control based on in-furnace temperature distribution. In com- ¼½ 1; 2; ...; K T parison to commercial and research projects described in [32–34], u u u u the innovative approach in this project relays on the on-line opti- T mization of fire-ball shape. By analyzing AGAM temperatures, SILO y ¼½y1; y2; ...; yR calculates current shape parameters of the fireball and compares T them with a reference fire-ball shape. The goal for SILO is to control z ¼½z1; z2; ...; zD the combustion process to minimize the difference between reference and current shape. This will positively influence the process The information about input-output relations is stored in parameters, related to boiler efficiency e.g. flue gas temperature, knowledge units – lymphocytes. Lymphocytes stores static excess air or CO emission. response of process outputs on a given change on process input signal, at certain process operating point. Process operating point 2. Coal combustion modelling is defined by disturbance signals. A single kth lymphocyte is presented below. In most optimization projects of the combustion process there ¼½k; k; D k; k; k; D k; ...; k are two main types of models – CFD and empirical models. CFD Lk ub ua u yb ya y z models are very complex and are used for deep investigation of the process. Information provided from CFD analysis could be used The model of the process in artificial immune system is calcu- for manual tuning of a boiler. The process of model creation, defi- lated basing on selected lymphocytes that represent current pro- nition of initial and boundary condition as well as calculations are cess operating point. It is a static and linear model. Process time consuming. That is why CFD could not be efficiently used for operating point is defined by the following vectors. closed-loop control of the combustion process. ¼½ a; a; a For the purpose of standard combustion optimization empirical A u y z models are used. The model consists of, but it is not limited to the The selection of the lymphocytes that fits to the operating point following input, output and disturbance signals: is govern by the following equation. ! ! d Inputs: YK YK ub k a ua k a lðLk; AÞ¼ g ðu ; u Þ g ðu ; u Þ – Secondary air dampers, i bi i i ai i ¼ ¼ – OFA dampers, i 1 i 1 – OFA tilts, ! ! ! YR YR YD – Coal feeders, gyb ðyk ; yaÞ gya ðyk ; yaÞ g zðzk; zaÞ j bj j j aj j l l l – Burner tilts, j¼1 j¼1 l¼1 – Oxygen setpoint bias, – ID and FD bias, The g functions are condition functions and examples of these d are presented below: Outputs: ( – Steam temperatures, jk aj > : z k a 0 if zl zl 0 01 – NOX emission, g ðz ; z Þ¼ l l l 1 if jzk zaj 6 0:01 – CO emission, l l

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 714 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720 8 < k > a > 0 if yb 10 or yj 10 gyb ðyk ; yaÞ¼ j j bj j : 1 if yk 6 10 or ya 6 10 bj j gub ðuk ; uaÞ¼1 i bi i Once the lymphocytes has been selected the algorithm calculates static gains that finally represent the relations in the process. The model is created every optimization step basing on information about current process operating point as well as using newest lymphocytes. During the operation, the artificial immune algorithm constantly records new lymphocytes that are utilized in next Fig. 1. An example of industrial application of AGAM system (Rybnik, unit 4). optimization steps. The new approach in combustion process modelling and opti- l2 mization to which this work refers, relays on artificial immune sys- T ¼ 2 tem method, but the difference is that the process output in- j R s furnace temperature distribution. This temperature distribution It was analyzed in [32], that the precision of temperature mea- is aggregated in three fire-ball shape categories: surement due to changes in combustion gas composition stays below 1.4% and the general error does not exceed 2%. Additionally, Left-right hot spot position, due to the fact that AGAM technology relay on contactless temper- intensity, ature measurement, radiation does not play a role on measure- dispersion. ment results – AGAM measures row gas temperature.

It is assumed and proved within this work that controlling the shape of fireball, combustion optimization is much more precisely 4. SILO – System for on-line optimization of the combustion and, finally, achieve better results than standard approach. More- process over, the fireball shape model in this idea is a compromise between CFD analysis and empirical models. From one hand the relation SILO is one of representatives of software solutions, which in between process inputs and shape categories has been defined control theory are called AC [9]. AC is a general name for group with empirical methods – artificial immune system. From the of solutions, mainly software. As the goal of base control systems other hand the temperature distribution is strongly related with is to facilitate basic operation by automation of the process, AC is CFD results, what has been proved in [38]. The analysis has shown aimed to optimize this process to meet particular performance that for the same boundary and initial conditions CFD simulation and economic objectives of unit operation. and AGAM system provide close temperature profiles and error SILO is an AC-class software solution, which is aimed to perform calculated for average temperature is at the level of 2.5–3%. automatic, on-line optimization of industrial processes – combustion in power boilers in particular. The SILO’s inspiration is an immune system of living creatures. This method is described in 3. Acoustic temperature measurement system – AGAM [31,39–42] but here will be briefly reminded. Additionally, the comparison between SILO and MPC approach can be found in [42]. AGAM is the acoustic system for in-furnace temperature distri- In general, SILO consists of two main, independent modules: bution measuring is one of these. The system measures tempera- Knowledge Gathering and Optimization. The Knowledge Gathering ture distribution of combustion gases on horizontal cross-section module is aimed to collect information about the process charac- of the furnace. AGAM technology is described in details in [2– teristic. It monitors process signals and identifies static relations 4,32,33].In[32,33] authors provide a brief description and focus between process controlled inputs – MV and outputs – CV at cer- on system’s features, which are important from combustion opti- tain operating point – constant DV. Each static input-output rela- mization point of view. tion, for certain process operating point is stored in SILO’s The physical principle behind acoustic technology is the rela- database. Using this knowledge for optimization purposes, SILO tion between sound speed in a gas, temperature and composition is able to calculate ad-hoc static characteristic of the process for of this gas. The following equation represents this relation: different operating points. rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Functionality of the Optimization module relay on constant cal- j B C ¼ T culating and updating MV signals. Calculated MV values are trans- M ferred to base control layer as new setpoints or corrections to setpoints for boiler controlled devices. When optimizing the pro- In industrial applications, the system consists of transmitters cess, SILO’s Optimization module monitors one internal parameter and receivers placed at the same level of the furnace. A distance – quality indicator. Formula of the quality indicator represents all between transmitter and receiver is a single measuring path. Set optimization goals and their priorities. The formula for calculating of transmitters and receivers creates a measuring mesh (combina- quality indicator in SILO is presented below. tion of multiple paths).

Fig. 1 presents an example conﬁguration of transmitters/ Xnm hi ¼ ðj ^c ~ c j lmÞ þ ððj ^c ~ c j smÞ Þ2 receivers and corresponding measuring mesh of the AGAM system J ak m k mk sk þ bk m k mk sk þ installed in Rybnik power plant. It consists of 8 transmitters/ k¼1 receivers, which creates 21 measuring paths. The system measures Xny hi þ ðj ^ ~ j lyÞ þ ððj ^ ~ j syÞ Þ2 temperature through each path basing on ‘‘travel-time” of sound ck y k yk sk þ dk y k yk sk þ ¼ impulse. A single transmitter generates speciﬁc sound’s impulse; k 1 all other receivers are ‘‘listening”. After receiving the impulse, Finally, the main goal of the Optimization module is the mini- system calculates the travel time. The temperature is calculated mization of the quality indicator. The module searches for such basing on the following equation: MV change (Dmd), which minimize the following formula:

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720 715 ( "# Xnm ðj ^c þ D d ~ c j lmÞ min ak m k m mk sk þ To optimize the efficiency of combustion process, it is neces- D d c d ~ c sm 2 sary to minimize heat losses and maximize useful heat outputs. m ¼ þb ððjm^ þ Dm m js Þ Þ 2k 1 k k k k 3þ9 Both components are defined by combustion parameters, which ly = Xny c ðjy^ þ DmdK y~ js Þ þ are measured on-line and listed in Table 1. In standard combus- þ 4 k k k k k þ 5 d sy 2 ; tion optimization projects signals from the table are controlled ¼ d ððjy^ þ Dm K y~ js Þ Þ k 1 k k k k k þ directly by the optimization software. In the new approach the SILO system controls the in-furnace temperature distribution. The Kk matrix represents static gains in mathematical model, which are automatically identified at each optimization period Proper temperature distribution means proper combustion, what basing on the most recent knowledge gathered in the optimizer’s finally has a positive effect on combustion parameters and boiler database [43–45]. efficiency. In order to perform on-line control of the temperature distribution in the combustion optimization configuration the following 5. New method in combustion optimization in existing power signals must be identified. plant d MV (Manipulated Variables) – 29 signals The integrated SILO-AGAM solution has been implemented in Oxygen stepoint (combustion air demand) – 1 signal; Rybnik Power Plant, unit 4. The Rybnik power station consists of Secondary air dampers – 12 signals; eight, bituminous coal fired units – OP-650 type boiler and 225 OFA II – 6 signals; MW turbine. The boiler is a pulverized coal fired boiler with steam OFAIII – 2 signals; drum and natural circulation. Maximum continuous rating is 650 Protective air dampers – 2 signals; tons of steam generation per hour. The designed superheated and Coal feeders – 6 signals. reheated steam temperature is 540 °C and pressure respectively d CV (Controlled Variables) – 12 signals 13.5 MPa and 2.3 MPa. AGAM temperatures The boiler is equipped with six pulverizers (A, B, C, D, E and F) d DV (Disturbance Variables) – 3 signals and twenty-four burners – each pulverizer supplies coal to four Unit load; burners. This is a wall (front wall) fired boiler with three burners’ Configuration of operating coal pulverizers; rows (Fig. 2). Depending on load demand some pulverizers are Estimated coal quality (coal quality value is estimates by turned off. dividing currently generated MWs by currently coal flow, The air/flue gas system consists of three FD fans and three ID represented by sum of coal feeders speed) fans. Two FD fans supplies primary air to the pulverizers and one Details of the fireball method as well as results of the project are FD fan – secondary air to the windbox. The secondary air is dis- presented in two following chapters. tributed through twelve secondary air dampers and sixteen OFAs. Each burner, in two lower rows has a dedicated secondary air dam- per. OFAs are arranged in two rows, OFA II installed on front wall and OFA III installed on rear wall. OFA I has been removed but 6. The fireball control algorithm the names remained unchanged. The boiler is equipped also with two protection air fans to protect the boiler’s evaporator against The main research goal for this project was to integrate and uti- corrosion. The AGAM system has been installed at furnace exit at lize in SILO the information about temperature distribution pro- 30.2 m level. vided by AGAM. In this chapter the research method and effects Regarding combustion control as well as optimization the boiler are described. efficiency is the main subject. Efficiency of power boilers could be Initially special test were performed to investigate whether calculated using two methods: input-output or heat loss [46]. The temperature distribution is controllable with boiler devices. Once heat loss method is mainly used for coal fired boilers and the effi- it had been confirmed further research has been carried on to ciency is calculated by the following formula: examine how a change in temperature distribution influences process outputs e.g. steam temperatures, NOX, CO, O2, etc. Detailed ¼ _ = _ ¼ _ =ð _ þ _ Þ gB 1 Q loss Q tot 1 Q loss Q out Q loss description of AGAM tests could be found in [47].

Fig. 2. Fuel and air supply system in Rybnik’s unit 4 boiler, where A1, A2, A3 and A4 are the burners supplied from pulverizer 1, B1, B2, B3 and B4 are the burners supplied from pulverizer 2, C1, C2, C3 and C4 are the burners supplied from pulverizer 3, D1, D2, D3 and D4 are the burners supplied from pulverizer 4, E1, E2, E3 and E4 are the burners supplied from pulverizer 5, F1, F2, F3 and F4 are the burners supplied from pulverizer 6.

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 716 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720

Table 1 Boiler parameters directly related to boiler’s efﬁciency.

Efficiency component Process parameters Useful heat output Superheated steam temperature, reheated steam temperature, reheated spray flow Heat loss in dry flue gas Mass of dry flue gas, flue gas temperature

Heat loss from H2O in combustion air, fuel Flue gas temperature and combustion products Heat loss from unburned carbon in the Mass of unburned carbon in refuse, fly and bottom ash Heat loss from unburned combustible gas in Mass of combustible gases in flue gases flue gases

Once relation between setpoints of MV signals, temperature Fig. 3. Penalty function for single AGAM temperature – standard approach. distribution and process outputs has been examinated, the categories of ﬁreball shape were deﬁned: d Left-right hot spot position, d intensity, d dispersion.

The fire-ball shape is calculated on a basis of twelve AGAM temperatures. Position of temperatures hot spot indicates, wheather the fire-ball is shifted to left or right side of the furnace. Intensity, is calculated as average temperature for a given temperature profile and dispersion is a parameter calculated as standard deviation of AGAM temperatures. After defining shape categories, historical values of AGAM temperatures as well as process parameters have been analyzed to find the best values of the fire-ball categories (reference shape), which represent most efficient combustion. To control the fireball shape in on-line mode it is necesarry to calculate adjustments to current shape. For this purpose gradient optimization algorithm was developed. This algorithm monitors Fig. 4. Penalty function for single AGAM temperature - dynamically calculated. difference between current and reference shape, and calculate optimal setpoints for twelve AGAM temperatures to minimize this tion of standard penalty function for single AGAM temperature and difference. The algorithm searches for such minimal change of cur- Fig. 4 presents new method – the penalty function for dynamically rent AGAM temperatures that finally gives reduced difference calculated setpoints. between current and reference shapes. For this particular example i.e. AGAM temperature – T , the The method used in gradient optimization is the unconstrained AGAM1 setpoint is 1350 °C, the linear insensibility zone (linear tolerance) gradient decent method. The objective function of this optimiza- equals ± 20 °C and the square insensibility zone (square tolerance) tion task consists of differences between values of current and ref- is ± 50 °C. The measured value – T , in this example, stays erence shape categories. The following formula represents this AGAM1 within linear tolerance, so penalty function returns zero – objective function: PAGAM1 = 0 (green dashed line). ! ! ! The result of gradient optimization is twelve optimized set- ¼j~ ð Þ^ jþj~ ð Þ^ jþj~ ð Þ^ j J sp T AGAM sp si T AGAM si sd T AGAM sd points of AGAM temperatures. Each single setpoint is uploaded The result of the gradient optimization is a vector of twelve to SILO as setpoint in SILO’s penalty function. Fig. 4 presents how AGAM temperatures. It is calculated by minimization of the objec- SILO interprets this setpoint’s change. As a result of gradient opti- tive function, by the following optimization task: mization, setpoint for this temperature changed – previously it was 1350 °C – Setpoint 1, the new equals 1380 °C – Setpoint 2. Linear ! ! ! ! ! and square insensibility zones stay unchanged and equal respec- ! ~ ^ ~ ^ ~ minDT jspðT AGAM þ DTÞspjþjsiðT AGAM þ DTÞsijþjsdðT AGAM tively ±20 °C and ±50 °C. After setpoint change, the TAGAM1 temper- ! ^ ature is now beyond linear tolerance range and, consequently, a þ DTÞsdj penalty is applied – PAGAM1 > 0 (blue dashed line). In this case, in AGAM temperatures are defined in SILO as regular CV signals, so next optimization step, SILO calculates change of MV signals, SILO collects the knowledge about relation between boiler con- which reduce this penalty and consequently, increase the trolled devices – MV and each single AGAM temperature. When, temperature. as a result of gradient optimization, setpoints of AGAM tempera- The following plots represent single step of the fireball opti- tures are changed, standard SILO algorithm utilize this knowledge mization algorithm. to calculate such change of MV signals, which minimize difference The first plot (Fig. 5) presents twelve AGAM temperatures mon- between the measured temperature and its new setpoint. itored by SILO – before optimization. In this case the fireball is Dynamically calculated setpoints for CV signals is a novel shifted to the right, rear side of the furnace. This shape differs from approach among all SILO implementations and refers to the defini- the reference significantly. Basing on difference between current tion of SILO’s objective function. Fig. 3 presents an example defini- and reference shape, the fire-ball optimization algorithm calculates

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720 717

Fig. 5. AGAM temperatures measured before SILO optimization.

Fig. 6. AGAM temperatures’ setpoints, calculated by gradient optimization.

Fig. 7. AGAM temperatures measured after SILO optimization.

Fig. 8. SILO effect on boiler efﬁciency.

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 718 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720 setpoints of AGAM temperatures. It gives twelve optimal setpoints 7. Optimization results of AGAM temperatures, presented on the second plot (Fig. 6). Then, those setpoints are transferred to SILO algorithm for further com- Evaluation of optimization results was done by analyzing his- bustion optimization. The ﬁnal result of this example, after SILO torical data of the combustion process parameters. The method optimization step, is presented in the third plot (Fig. 7). of this evaluation relays on simple compering, how SILO operation

Fig. 9. SILO effect on left and right SH steam temperatures.

Fig. 10. SILO effect on left and right RH steam temperatures.

Fig. 11. SILO effect on difference of RH steam temperature.

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720 719

Fig. 12. SILO effect on aggregated boiler efficiency. influences individual CV signals and efficiency of the process. In Protection and Water Management [grant number: GEKON1/ other words, the difference in process parameters and efficiency O2/213655/9/2014]. between SILO ON and SILO OFF states will be presented. The data- set represents the period of August 1st – December 31st 2014. References Before calculating final results, two statistical analyses must be performed – Kolmogorov-Smirnov [48] as well as Mann-Whitney- [1] Key World Energy Statistics 2015, International Energy Agency, Paris, Wilcoxon [49] tests. Those tests help to evaluate, whether a change November 2015. [2] D. Husson, S.D. Bennett, G.S. Kino, Remote temperature measurement using an of particular parameter is caused by certain, intended impulse e.g. acoustic probe, Appl. Phys. Lett. 41 (1982) 915, http://dx.doi.org/10.1063/ the gain in boiler efficiency was the effect of SILO operation. 1.93334. Detailed information of optimization results are presented on [3] Hua Yan, Zhen Peng, Kexin Cui, Liang Zhang, Acoustic travel-time Figs. 8–11. measurement in acoustic temperature field monitoring, in: 2008 7th World Congress on Intelligent Control and Automation, Chongqing, China, June 2008, Fig. 8 presents positive influence of SILO on the boiler efficiency pp. 4947–4951. for the whole load range. The highest increase was recorded for [4] Kousuke Kudo, Koichi Mizutani, Temperature measurement using acoustic higher load range and the lowest for low load. reflectors, Jpn. J. Appl. Phys. 43 (5B) (2004) 3095–3098. [5] Hajime Arimoto, Nobuo Takeuchi, Sachio Mukaihara, Toru Kimura, Ryuzo Result of SILO operation on superheated and reheated steam Kano, Takeo Ohira, Shinji Kawashima, Kazuya Iwakura, Applicability of TDLAS temperatures represents highest optimization priority on those gas detection technique to combustion control and emission monitoring under signals. Steam temperatures influence boiler efficiency signifi- harsh environment, Int. J. Technol. 2 (1) (2011) 1–9. [6] Pei-jin Liu, Bin Huang, Bin Yang, Guo-qiang He, TDLAS for measurement of cantly, because they influence useful heat output of the boiler [46]. temperature in combustion environment, in: Proceedings of SPIE - The Regarding reheat steam temperature, the optimization effect is International Society for Optical Engineering, 2013, vol. 8796; http://dx.doi. mostly seen on the left-side temperature at low and medium unit org/10.1117/12.2011262. [7] D. Nabagło, P. Madejski, Combustion process analysis in boiler OP-650K based load. For high load level this temperature stays within the toler- on acoustic gas temperature measuring system, in: 3rd International ance – ±5 °C around 540 °C. Conference on Contemporary Problems of Thermal Engineering CPOTE 2012, Fig. 11 presents an improvement on left-right difference of 18–20 September 2012, Gliwice, Poland. [8] Andrew D. Sappey, Pat Masterson, Eric Huelson, Jim Howell, Mike Estes, Henrik reheated steam temperatures. This parameter was significantly Hofvander, Atilio Jobson, Results of closed-loop coal-fired boiler operation reduced for low load range. For other operating points this problem using a TDLAS sensor and smart process control software, Combust. Sci. did not occur – the parameter stayed within tolerance range. Technol. 183 (11) (2011) 1282–1295. [9] P. Tatjewski, Advanced Control of Industrial Processes: Structures and Aglorithms, Springer Verlag, London, 2007. 8. Conclusions [10] E.F. Camacho, C. Bordons, Model Predictive Control, Springer Verlag, London, 1999. [11] Havlena Vladimı´ r, Findejs Jirˇ´ı; Application of model predictive control to The Rybnik’s project has a positive influence on the combustion advanced combustion control, Control Eng. Practice 13(6) (2005) 671–680. process. The goal of this project was to increase the process effi- [12] J. Arabas, L. Białobrzeski, T. Chomiak, P.D. Doman´ ski, K. S´wirski, R. ciency over 0.2%. As it is presented below, final average efficiency Neelakantan, Pulverized coal fired boiler optimization and NOx control using neural networks and fuzzy logic, in: Proc. of AspenWorld’97, Boston, increase is 0.27% (see Fig. 12). Massachusetts, October 1997. By monitoring temperature distribution, SILO is able to control [13] J. Arabas, L. Bialobrzeski, P.D. Doman´ ski, K. S´wirski, Advanced boiler control, the combustion process to be more balanced. Balanced tempera- in: Proc. of MMAR’98, Miedzyzdroje, Poland, 1998. [14] Erik Schaffernicht, Volker Stephan, Klaus Debes, Horst-Michael Gross, Machine ture distribution means balanced O2 distribution and, conse- Learning Techniques for Selforganizing Combustion Control Chapter, Lecture quently, homogeneous combustion. This, finally, results in lower Notes in Computer Science 5803 (2009) 395–402. CO emission – due to reduction of local under-stoichiometric com- [15] Soteris A. Kalogirou, Artificial intelligence for the modeling and control of combustion processes: a review, Prog. Energy Combust. Sci. 29 (2003) 515– bustion. This opens for SILO a potential of further decreasing 566. demand for combustion air. [16] Mariusz Kalita, Waldemar Wójcik, Andrzej Smolarz, Conception of genetic A homogeneous temperature distribution is a great indicator of controller application in power boiler, Proc. of SPIE 5948 (2005). [17] P. Ilamathi, V. Selladurai, K. Balamurugan, V.T. Sathyanathan, ANN-GA properly controlled combustion process. Due to the characteristics approach for predictive modeling and optimization of NOx emission in a of this process, distribution of fuel and air must be adjusted contin- tangentially fired boiler, Clean Techn. Environ. Policy 15 (2013) 125–131. uously, in order to keep it at high quality. [18] W. Wojcik, M. Kalita, A. Smolarz, B. Pilek, R. Romaniuk, Controlling combustion process in power boiler by genetic algorithm and neural network, in: Photonics Applications in Astronomy, Communications, Industry, and High- Acknowledgments Energy Physics Experiments III, May 5775(1) (2004) 348–353. [19] Jussi Makila, Juha-Pekka Jalkanen, Neural Network Combustion Optimisation in Naantali Power Plant, Artificial Neural Nets and Genetic Algorithms, in: Funding: This research was funded by the National Center of Proceedings of the International Conference in Prague, Czech Republic, 2001 Research and Development and National Fund for Environmental Pages, pp. 185–188.

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24 720 Ł. S´ladewski et al. / Applied Thermal Engineering 123 (2017) 711–720

[20] Ilamathi Balamurugan, Selladurai V. Gounder, Balamurugan Kulendran, ANN - operation using a TDLAS sensor and smart process control software, SQP Approach for NOx Emission Reduction In Coal Fired Boilers; Int. J. Emerg. Combust. Sci. Technol. 183 (11) (2011) 1282–1295. Elect. Power Syst. 13(3) (2012). [36] Benyuan Huang, Zixue Luo, Huaichun Zhou, Optimization of combustion based [21] Guoqiang Li, Peifeng Niu, Combustion optimization of a coal-fired on introducing radiant energy signal in pulverized coal-fired boiler, Fuel boiler with double linear fast learning network, Soft. Comput. 20 (1) (2016) Process. Technol. 91 (2010) 660–668. 149–156. [37] A. Rahrooh, S. Shepard, Identification of nonlinear systems using NARMAX [22] Huan Zhao, Pei-Hong Wang; Modeling and optimization of efficiency and NOx model, Nonlin. Anal.: Theory, Meth. Applicat. 71 (12) (2009) e1198–e1202. emission at a coal-fired utility boiler; 2009 Asia-Pacific Power and Energy [38] N. Modlinski, P. Madejski, T. Janda, K. Szczepanek, W. Kordylewski, A Engineering Conference, March 2009, pp. 1–4. validation of computational fluid dynamics temperature distribution [23] A.J. Smola, B. Scholkopf, A tutorial on support vector regression, Stat. Comput. prediction in a pulverized coal boiler with acoustic temperature 14 (2004) 199–222. measurement, Energy 92 (2015) 77–86. [24] Wang Weiqing, Multi-objective Optimization of Coal-Fired Boiler Efficiency [39] K. Wojdan, System of optimizing steady-state control of technical processes and NOx Emission under Different Ecological Environment; Future inspired by operation of an immune system, Warsaw University of Communication, Computing, Control and Management, Springer Berlin Technology, PhD thesis, Warsaw, 2008. Heidelberg, 2012, pp. 433–439. [40] K. Wojdan, K. Swirski, M. Warchol, Transition States Handling in Self-Adaptive [25] Zhongbao Wei, Xiaolu Li, Lijun Xu, Yanting Cheng, Comparative study of Steady State Optimizer of Industrial Processes, in: Proceedings of the IASTED computational intelligence approaches for NOx reduction of coal-fired boiler, Conference on Modelling, Identification, and Control, Thailand, Phuket, ACTA Energy 55 (2013) 683–692. Press, 2010. [26] Li-Gang Zheng, Hao Zhou, Ke-Fa Cen, Chun-Lin Wang, A comparative study of [41] K. Wojdan, K. S´wirski, T. Chomiak, Immune Inspired System for Chemical optimization algorithms for low NOx combustion modification at a coal-fired Process Optimization using the example of a Combustion Process in a Power utility boiler, Expert Syst. Appl. 36 (2) (2009) 2780–2793. Boiler, in: IEEE Proc. of 14th International Conference on Intelligent System [27] L.G. Zheng, L.H. Jiang, M.G. Yu, Support vector regression and ant colony Applications to Power Systems -ISAP 2007, Kaohsiung, Tajwan, 2007. optimization for pollutant emission control in power plants, Prog. Environ. Sci. [42] K. Wojdan, K. S´wirski, M. Warchol, Transition state layer in the immune Technol. 1 (2007) 550–554. inspired optimizer; Trends in applied artificial intelligence, Lecture Notes in [28] Feng Wu, Hao Zhou, Jia-Pei Zhao, Ke-Fa Cen, A comparative study of the multi- Artific. Intell. 6096 (2010) 11–20. objective optimization algorithms for coal-fired boilers, Expert Syst. Appl. 38 [43] K. Wojdan, K. S´wirski, M. Warchoł, M. Maciorowski, Conditioning of Model (2011) 7179–7185. Identification Task in Immune Inspired Optimizer SILO; Chapter in book: [29] Jianhua Zhangl, Bin Tian, Guolian Houl, Jinfang Zhangl, Improvement of Boiler IAENG Transactions on Engineering Technologies Volume 3, American Combustion Control Performance Using Probability Density Function Shaping Institute of Physics (AIP), listopad 2009. and Particle Swarm Optimization, in: 2008 2nd International Symposium on [44] K. Wojdan, K. S´wirski, M. Warchoł, M. Maciorowski, Maintaining good Systems and Control in Aerospace and Astronautics, December 2008, pp. 1–5. conditioning of model identification task in immune inspired on-line [30] Xingrang Liu, R.C. Bansal, Integrating multi-objective optimization with optimizer of an industrial process, Engineering Letters, wolumen 17 (2) computational fluid dynamics to optimize boiler combustion process of a (2009) 93–100. coal fired power plant, Appl. Energy 130 (2014) 658–669. [45] K. Wojdan, K. S´wirski, M. Warchoł, M. Maciorowski, Methods providing good [31] K. Wojdan, K. S´wirski, M. Warchol, J. Milewski, A. Miller, A practical approach conditioning of model identification task in immune inspired, steady-state to combustion process optimization using an improved immune optimizer, controller of an industrial process, in: Proc. Of International MultiConference Sustain. Res. Innovat. Proc., vol. 3, Kenya, 2011. of Engineers and Computer Scientists 2009, vol. II, IMECS 2009, Hong Kong, [32] M. Deuster, Acoustic gas temperature measurement, in: Proceedings of 2009. Wissenforum: temperature measurement technique Conference, [46] Water tube boilers and auxiliary installations – part 15: acceptance tests, EN Aldenhoven, 2009. 12952–15, European Committee for Standardization. [33] M. Deuster, Mit Schallgeschwindigkeit berührungslos hohe Gastemperaturen [47] Ł. S´ladewski, D. Nabagło, T. Janda, J. Chachuła, Combustion process messen, Sonderdruck MSR Magazin, 1998. optimization by using immune optimizer in power boiler, Archivum [34] E. Huelson, N. Logan, A.D. Sappey, G. Tanck, Ch. Steiger, N. Jakinovich, J.P. Scott, Combustionis, 35(1) (2015). T. Alleshouse, P. Spinney, J. Grott, H. Winn, Carbon Management for Existing [48] MIT open courseware, ‘‘Kolmogorov-Smirnov test”, w Statistics for Power Plants via Measurement and Control Optimization, DOE/NETL-2011, Applications Fall Lecture, Massachusetts Institute of Technology, January 28, 2011. Massachusetts, 2006, pp. 83–90. [35] Andrew D. Sappey, Pat Masterson, Eric Huelson, Jim Howell, Mike Estes, [49] N. Nachar, The Mann-Whitney U; Tutorials in Quantitative Methods for Henrik Hofvander, Atilio Jobson, Results of closed-loop coal-fired boiler Psychology, Université de Montréal, Montréal, 2008, pp. 13–20.

Pobrano z http://repo.pw.edu.pl / Downloaded from Repository of Warsaw University of Technology 2021-09-24