Superheater Steam Temperature Control for a 300MW Boiler Unit with Inverse Dynamic Process Models

1 Superheater Steam Temperature Control for a 300MW Boiler Unit with Inverse Dynamic Process Models

Liangyu Ma, Yongjun Lin and Kwang Y. Lee, Life Fellow, IEEE

to achieve good control results under different loading Abstract--An Inverse Dynamic Neuro-Controller (IDNC) is condition and changing environment, which often requires developed to improve the superheater steam temperature control much effort and time [1]. Therefore, it is highly desirable to of a 300MW boiler unit. A recurrent neural network was used adopt intelligent control algorithms in order to improve the for building the Inverse Dynamic Process Models (IDPMs) for superheater steam temperature control. the superheater system. Two inverse dynamic neural network (NN) models referring to the first-stage and the second-stage Artificial neural network is an attractive method for water-spray attemperators are constructed separately. To identifying nonlinear processes, due to its good modeling achieve highly accurate approximation of the superheater system, capability and its ability to learn complex dynamic behaviour the NN models are trained with sufficient historical data in a of a physical system. Applications of neural networks have wide operating range, which consists of both different steady- been widespread in process control, both in simulation and state conditions and dynamic transients. Then the IDNCs are on-line implementation, which includes model predictive designed based on the well-trained IDPMs and applied to superheater steam temperature control. In order to eliminate the control, inverse model-based control, and adaptive control [2]. steady-state control error arisen by the model error, a simple Recently, the concept of inverse dynamic neuro-controller was feedback PID compensator is added to an inverse controller. used for superheater and reheater steam temperature control of Detailed control tests are carried out on a full-scope simulator for a large-scale ultra-supercritical boiler unit [3,4]. Feedforward a 300MW coal-fired power generating unit. It is shown that the and recurrent networks are the two commonly used neural temperature control is greatly improved with the IDNCs network structures for modeling, prediction and control of compared to the original cascaded PID control scheme. nonlinear dynamic systems. A feedforward network is a static Index Terms—Boiler control, feedforward control, inverse mapping that models a steady-state condition of a plant. It can dynamic neuro-controller, superheater steam temperature also be used to model dynamic behavior of the plant by control, neural network control, PID compensator. including delayed input and output values as additional inputs. On the other hand, a recurrent network possesses internal I. INTRODUCTION memory by including feedbacks for past values, either from UPERHEATER steam temperature is one of the most network output into hidden units or from hidden units in one Simportant variables for a boiler system in a large-capacity layer into hidden units in another layer [5-8]. fossil-fuel power generating unit. It must be controlled to be The main focus of this paper is to use a recurrent neural within given upper and lower limits to ensure high efficiency network to model the inverse dynamic relationship of the and safety of the power unit. However, often it is not easily superheater steam temperature in a 300MW boiler unit and controlled with a simple PID controller since the boiler is a use the trained inverse dynamic neural network models as multi-input multi-output (MIMO) nonlinear system consisting feedforward controllers to improve the superheater steam of many strongly-coupled sub-systems, which leads to a large temperature control. Detailed control tests are made on a full- time delay and big inertia to the superheater steam scope simulator of a 300MW power generating unit. temperature response. Therefore, cascaded PID controllers are generally used for superheater steam temperature control and II. SYSTEM DESCRIPTION the control logic is getting more complex with increasing The power plant under investigation is a 300MW boiler capacity. Moreover, even for a cascaded control scheme subcritical coal-fired boiler-turbine-generator unit. The rated with at least 2 PID control modules, the gains and time steam temperature for superheater and reheater under full load constants of these PID controllers have to be tuned frequently condition is 540℃. The superheater is mainly made of 4 parts: (1) Primary Superheater (PSH), which is arranged above the L. Y. Ma and Y. J. Lin are with the Automation Department, School of economizer, including both horizontal-type and vertical-type Control and Computer Engineering, North China Electric Power University, ones; (2) Divided Platen Superheater (DPSH), which is Baoding, Hebei 071003, China (e-mail: [email protected]; [email protected]). installed over furnace combustion area and between the front K. Y. Lee is with the Department of Electrical and Computer Engineering, waterwall and the Rear Platen Superheater; (3) Rear Platen Baylor University, Waco, TX 76798-7356, USA (e-mail: Superheater (RPSH) located before the furnace arch; and (4) [email protected]). Final Superhater (FSH) installed at the horizontal gas pass.

The reheater also includes 3 different parts: (1) Wall-type III. NEURAL NETWORK STRUCTURE AND ITS TRAINING Radiant Reheater, which is arranged near the front and side- SCHEME front waterwall, about 1.5m lower than the DPSH; (2) Rear Recurrent neural network differs from other conventional Platen Reheater (PRH), installed behind the RPSH along the feedforward networks in that it includes recurrent or feedback gas direction; (3) Final Reheater (FRH), installed in the connections [6-8]. The delays in these connections store horizontal gas pass behind the furnace arch and before the values from the previous time step, which makes it sensitive to FSH. the history of input and output data and fit for dynamic system During normal operation, the superheater steam modeling. For convenience, the Elman network is often used temperature is adjusted by regulating the first stage and the for a recurrent neural network, which has tansig neurons in its second stage water-spray attemperators. The first stage hidden (recurrent) layer, and purelin neurons in its output attemperator is installed between the PSH and the DPSH. The layer [9-11]. This combination is special in that a three-layer second stage attemperators are installed between the RPSH network with these transfer functions can approximate any and FSH. The desuperheating water comes from the feedwater function (with a finite number of discontinuities) with pumps’ outlet header. As for the reheater, the reheated steam arbitrary accuracy if the hidden layer has enough neurons. The temperature is mainly controlled by adjusting the angle of the structure of an Elman recurrent network is shown in Fig. 3. tilting burners, either upward or downward under normal operation. The angle adjustment for the tilting burners also exerts influence on the superheater steam temperature. In addition, two paratactic water-spray attemperators are installed at the entrance of the wall-type radiant reheater, mainly for emergency desuperheating use. The reheater desuperheating water comes from the midpoint taps of the running feedwater pumps. The schematic of the above 300MW boiler system is shown in Fig. 1. The modularized flow chart of the superheater system is shown in Fig. 2.

Fig. 3. Structure of the Elman recurrent network model.

As shown in Fig. 3, the outputs in each layer of an Elman network is given by: MR =+ (1) x jijiiji()kfWukWock (∑∑ 1,, () 3 _ ()) ii==11 =− ckii() xk ( 1) (2)

R = y jiji()kgWxk (∑ 2, ()) (3) i=1 where, is the weight that connects node i in the input W1i, j

layer to node j in the hidden layer; W 2i, j is the weight that connects node i in the hidden layer to node j in the output layer; is the weight that connects node i in the context ①Economizer；②Steam Drum；③ Primary SH；④Divided platen SH；⑤ W 3i, j Rear Platen SH；⑥Final SH；⑦wall-type radiant reheater；⑧Rear Platen layer to node j in the hidden layer; and f(·) and g(·) are the RH；⑨Final RH transfer functions of hidden layer and output layer, Fig. 1. Schematic of a 300MW boiler unit. respectively. An Elman neural network can be created and trained according to the back-propagation algorithm with MATLAB Neural Network Toolbox. When the entire input sequence is presented to the network, its outputs are calculated and compared with the target sequence to generate an error sequence. For each time step, the error is back propagated to find gradients of errors for each weight and bias. This gradient is actually an approximation, because the contributions of Fig. 2. Modularized structure of the superheater system. weights and biases to errors via the delayed recurrent 3 connection are ignored [9-11]. However, more accurate gradient can be evaluated by including the contributions Since an inverse dynamic process model expressed by the through the recurrent neurons [8]. This gradient is then used to neural network is not a complete inverse but an update the weights with the chosen back-propagation training approximation, it may generate a steady-state error when it is algorithm [12]. Since Levenberg-Marquart method is fast and used as a feedforward controller. Therefore, a supplementary signal E(k) is needed to eliminate the steady-state error has robust convergence property in the off-line training, it is induced by modeling error and other disturbances. The used for training the Elman network in this paper. supplementary signal E(k) may come from the output of a feedback PID compensator. This combined control scheme IV. INVERSE CONTROL SCHEME with an IDPM feedforward controller and a simple PID The concept of inverse control was originated by B. feedback compensator is shown in Fig. 6. Widrow [13-16], and has been widely used in control of different industrial applications [3,17-21]. The inverse controller itself is an approximation of the inverse dynamic performance of the plant model. As stated above, good properties of a recurrent neural network make it competent in modeling the inverse dynamic performance of a complex system. The inputs and outputs of a NN-based Inverse Dynamic Process model (IDPM) for a given system can be determined by analyzing the system and problem carefully based on basic inverse control principle. Then the IDPM can Fig. 6. Inverse control scheme augmented by a PID compensator. be constructed and trained with enough historical inputs/outputs data sequence. V. INVERSE DYNAMIC PROCESS MODEL DESIGN FOR There are different methods for training the IDPM model. SUPERHEATER STEAM TEMPERATURE CONTROL The most commonly used one is called generalized learning method [18,19], as shown in Fig. 4. With this method, the A. Inverse Dynamic Process Model input of the IDPM is the plant output and the target IDPM By isolating the superheater system from the whole boiler output is the plant input. The training is executed off-line with unit and carefully analyzing the most important peripheral a considerable amount of output-input pairs to guarantee the variables influencing the superheater steam temperature, it is neural network IDPM capturing the total inverse dynamics of found that the superheater system is a very complex MIMO the plant. system. Many variables, such as air flow, feedwater flow, coal flow, burner tilting angle and several water-spray attemperators, influence the superheater steam temperature. Therefore, all these factors should be considered for designing the inverse dynamic process model. According to the original boiler control logic, spray-1 and spray-2 are used for controlling the outlet temperatures of Divided Platen SH and the Final SH, respectively. Therefore, two similar neural network IDPMs are designed for superheater steam temperature control. One model deals with the first stage water-spray attemperator and the Divided Platen Fig. 4. Training of the IDPM. SH system, which will be used to regulate the outlet After the training is finished with sufficient accuracy, the temperature of the Divided Platen SH. Another model is inverse model can be used as a feedforward controller by related to the second stage attemperators and the Final SH replacing the input of the inverse model with the expected system, which will be used to control the final SH outlet steam setpoint y (k). How the IDPM is used as a feedforward temperature. Each neural network inverse model includes 1 ref output, i.e., spray-1 or spray-2 control valve opening demand controller is shown in Fig. 5. If the network represents the (Since the left-hand side and right-hand side spray-2 control exact inverse, the control output ur(k) produced by the valves usually keep the same opening under auto mode, it is network will drive the system future output y(k +1) simplified as one spray-2 control valve in our work). The to y (k). platen SH outlet steam temperature and the final SH outlet ref steam temperature are among the total 8 inputs of respective model. The inputs and output of the 2 neural network IDPMs are shown in Tables I and II.

B. Training Data Preparation Fig. 5. The IDPM used as a feedforward controller. An inverse neural network controller can be viewed as a 4 feedforward controller. Its main function is to provide fast The outputs of the trained neural networks are compared control when the unit load demand is changed. Thus, it is with the actual control demands (data used for training) in essential to have data for a wide-range operating conditions, Figs. 7 and 8, where dashed lines are for the outputs of the both steady-state and dynamic transients, in order to train the neural networks. The figures show that the two IDPMs are neural networks to ensure that the neural network models trained with sufficient accuracy and are ready for use as capture the comprehensive inverse dynamics of the system. If feedforward controllers. the data used for network training is not sufficient an inverse 35 controller cannot be counted in giving reliable control demand under different operating conditions. Therefore, data selection 30 is a very important factor for an IDPM development. 25 In this paper, the following operating conditions are included in the training data, totaling 2990 epochs: (1) 20 different steady-state conditions, 100%, 90%, …, 70% load 15 levels; (2) load change from 100 to 90% load levels; (3) load 10 change from 90 to 80%; and (4) load change from 80 to 70%. 5

Spray-1control demand % Actual output TABLE I. 0 INPUT/OUTPUT VARIABLES OF THE IDPM FOR SPRAY-1 CONTROL. NN model output

-5 0 500 1000 1500 2000 2500 3000 Inputs (8) Epoch (1) Coal flow (Kg/h) Fig. 7. Spray-1 control valve opening. (2) Air flow (Km^3/h)

(3) Feedwater flow (Kg/h) 50 (4) Feedwater pressure (Mpa) 45 (5) Burner tilting angle（°）

(6) Main steam pressure （Mpa） 40 (7) Primary SH Outlet steam temperature（℃） 35 (8) Divided Platen SH outlet steam temperature (℃)

Output (1) 30 (1) Spray-1 control valve opening (%) Spray-2 control demand % 25 Actual output TABLE II. NN model output INPUT/OUTPUT VARIABLES OF THE IDPM FOR SPRAY-2 CONTROL 20 0 500 1000 1500 2000 2500 3000 . Epoch Inputs (8) Fig. 8. Spray-2 control valve opening. (1) Coal flow (Kg/h) (2) Air flow (Km^3/h) VI. CONTROL SIMULATION TESTS (3) Feedwater flow (Kg/h) After the off-line training has been finished with sufficient (4) Feedwater pressure (Mpa) accuracy, the two IDPMs are used as feedforward controllers (5) Burner tilting angle（°） for spary-1 and spray-2 control valves. For the feedforward (6) Main steam pressure （Mpa） control action to take place, an input to the IDPM needs to be replaced by a desired or reference output as illustrated in Fig. (7) Rear Platen SH outlet steam temperature（℃） 5. The setpoint for spray-1 is the Divided Platen SH outlet (8) Final SH outlet steam temperature (℃) temperature, which is the model’s 8th input in Table I. Output (1) Similarly, the setpoint for spray-2 is the Final SH outlet steam (1) Spray-2 control valve opening (%) temperature, which is the model’s 8th input in Table II. All other 7 inputs for both IDPMs are the endogenous inputs C. Training and Development of the IDPM coming from real-time simulation of the plant. For the two IDPM neural networks, each with 8 inputs and For the 300MW boiler unit, the setpoint of the Divided 1 output, their optimal hidden neuron numbers are determined Platen SH outlet temperature T1sp is fixed at 470℃, and the with a MATLAB optimal search program: 14 and 12 for the setpoint of the Final SH outlet steam temperature T2sp is fixed first and second IDPMs, respectively. Then the two networks at 540 ℃ . The SH outlet steam temperatures cannot be are trained and after 75 epochs of training, the mean-squared changed instantaneously when the difference between the error (MSE) for the first IDPM is 4.323e-06, and that for the setpoint and current temperature is big. Thus, reference SH second IDPM is 9.364e-8. outlet temperatures for the IDPMs, T1ref(k) and T2ref(k), are 5 adjusted by:

T1ref(k)= T1(k)+sat[T1sp-T1(k)] (4)

T2ref(k)= T2(k)+sat[T2sp-T2(k)] (5) where, T1(k),and T2(k) are the outlet temperatures at time k for the Divided Platen SH and the Final SH, respectively, and sat[.] is the saturation function, which limit the change of the reference value to be within the ramp rate. With above scheme, detailed control tests have been carried out on a full-scope simulator of the 300MW subcritical boiler-turbine-generator unit. The tests included 3 cases: conventional cascaded control, inverse control, and inverse control with a simple PI compensator augmenting the second water-spray control valve as shown in Fig. 6. The inverse control program was developed with MATLAB. The real-time (c) Load drop from 240MW to 210MW, rate 5MW/min data communication between MATLAB and the simulator Fig. 9. Comparison of the results with different control schemes: (1) was accomplished with UDP/IP protocol. Cascaded PID control, (2) NN inverse dynamic control, and (3) NN inverse For each test, three steps are included: (1) Change the unit control with PI compensator. load demand from 300MW to 270MW with a rate of It can be seen very clearly from these figures, the direct 5MW/min under sliding-pressure mode; (2) When the unit, inverse control has faster convergence speed and less overshot especially the Final SH outlet steam temperature, has reached than the conventional cascaded PID control scheme during the a stable state, continue to drop the load from 270MW to whole load-dropping processes (from 100% to 70% full load), 240MW; and (3) When the unit is stable again, continue to although the inverse dynamic process models are trained with drop the load from 240MW to 210MW. Comparisons for the the operating data under the conventional cascaded control. 3 different control schemes are shown in Fig. 9. Another point noticed is that the inverse control has generated a little steady-state control error. The error accumulated to 0.65 ℃ when the load is 210MW. This error is unavoidable since the inverse control scheme shown in Fig. 5 is in fact a feedfoward controller without any real-time feedback compensation. In order to eliminate the steady-state control error arisen by the IDPM control, the control scheme with a simple feedback PI compensator is introduced as shown in Fig. 6. A PI feedback controller was added only to the output of the second inverse controller. As shown by the curves marked with (3) in Figs. 9(a), (b) and (c), it is surprising to notice that with this additional PI compensator not only the steady-state error led by the inverse control was eliminated, but also the

overshot of the steam temperature was greatly reduced during (a) Load drop from 300MW to 270MW, rate 5MW/min the whole loading-down process from 300MW to 210MW.

VII. CONCLUSION In this paper, the concept of recurrent neural network based inverse dynamic process model (IDPM) is developed and used as feedforward controllers to improve the superheater steam temperature control for a 300MW boiler unit. The main purpose of an inverse dynamic neural controller is to shorten the stabilization time and reduce the overshot of the control process. In order to eliminate the steady-state control error induced by the IDPM, a simple feedback PI compensator is added to the inverse controller for the second spray. Simulation control tests on a full-scope simulator of the 300MW power generating unit have demonstrated the validity of the proposed control scheme in improving the superheater (b) Load drop from 270MW to 240MW, rate 5MW/min steam temperature control. 6

Since the superheater system of a large boiler is a very [20] A. Malinowski, J. M. Zurada, and J. H. Lilly "Inverse control of complex nonlinear MIMO system, the superheater steam nonlinear systems using neural network observer and inverse mapping approach," Proc. of IEEE International Conference on Neural Networks, temperature is influenced by many factors. The selection of Perth, Western Australia, 1995, vol. 5, pp. 2513-2518. the input variables for the IDPM model and the selection of [21] J. H. Zhang, G. L. Hou, and J. F. Zhang, "Adaptive neuro-control the data for model training both influence the control effect system for superheated steam temperature of power plant over wide range operation," Sixth International Conference on Intelligent Systems greatly. Other IDPM structures and different input Design and Applications (ISDA'06), 2006, pp. 138-141. combinations will be investigated and tested to achieve better control results in future.

IX. BIOGRAPHIES VIII. REFERENCES

[1] I. Benyo, Cascade generalized predictive control-applications in power Liangyu Ma received his B.S. degree in plant control. Oulu University Press, Finland, 2006. Thermal Power Engineering in 1993 from North [2] H. M. Azlan, "Review of the applications of neural networks in chemical China Electric Power University (NCEPU), process control—simulation and online implementation," Artificial Baoding, China. He received his M.S. and Ph. Intelligence in Engineering, 1999, no.13, pp. 55-68. D. degrees in 1996 and 2004, respectively, from [3] K. Y. Lee, L.Y. Ma, C. J. Boo, W.-H. Jung, and S.-Ho Kim, "Inverase NCEPU. dynamic neuro-controller for superheater steam temperature control of a Now he works as an associate professor with large-scale ultra-supercritical (USC) boiler unit," Proc. of the IFAC Automation Department, School of Control and Symposium on Power Plants and Power Systems Control, in Tampere, Computer Engineering, NCEPU. He was a Finland, July 5-8, 2009. visiting scholar at Baylor University, Waco, [4] K. Y. Lee, L.Y. Ma, C. J. Boo, W.-H. Jung, and S.-Ho Kim, "Intelligent Texas, USA, from March 2008 to February modified predictive optimal control of reheater steam temperature in a 2009. His special fields of research interest include power plant modeling large-scale boiler unit," Proc. of the IEEE Power and Energy Society and simulation, condition monitoring and intelligent fault diagnosis for General Meeting, in Calgary, Canada, July 26-30, 2009. power station thermal facilities and systems, intelligent control and [5] G. Irwin, M. Brown, B. Hogg, and E. Swidenbank, " Neural network application to power plants. modelling of a 200MW boiler system," IEE Proceedings-Control theory and Applications, 1995, 142(6), pp. 529-536. [6] R. Gencay and T. Liu. "Nonlinear modeling and prediction with feedforward and recurrent networks," 1997, Physica D 108, pp. 119-134. Yongjun Lin received his B. S. and M.S. [7] C.-C. Ku, K. Y. Lee, and R. M. Edwards. "Improved nuclear reactor degrees in Automation, and his Ph. D degree in temperature control using diagonal recurrent neural networks," IEEE Thermal Power Engineering, in 1986, 1989 and Trans. Nuclear Science, 1992, 39(6), pp. 2298-2308. 2002, respectively, from North China Electric [8] C.-C. Ku and K. Y. Lee. "Diagonal recurrent neural network for Power University (NCEPU), Baoding, China. dynamic systems control," IEEE Trans. Neural Networks, 1995, 6(1), pp. He is a professor with Automation 144-156. Department, School of Control and Computer [9] J. Elman, "Finding structure in time," Cognitive Science, 1990, vol. 14, Engineering, NCEPU. His main fields of pp. 179-211. research interest include application of [10] X. Z. Gao, X. M. Gao, and S. J. Ovaska. "A modified Elman neural advanced control strategies in power plant network model with application to dynamic systems Identification," control, embedded operating system and Proceedings of IEEE International Conference on Systems, Man and computer network communication and safety. Cybernetics, Beijing, China, 1996, pp.1376-1381. [11] Y. C. Cheng, W. M. Qi, and W. Y. Cai, "Dynamic properties of Elman and modified Elman neural networks," Proceeding of the 1st International Conference on Machine Learning and Cybernetics, Kwang Y. Lee received his B.S. degree in Beijing, 2002, pp. 637-640. TM Electrical Engineering in 1964 from Seoul [12] H. Demuth, M. Beale, and M. Hagan. Neural Network Toolbox 6 National University, Seoul, Korea, M.S. degree User’s Guide. The Math Works, Inc., 1997. in Electrical Engineering in 1968 from North [13] B. Widrow, J. McCool, and B Medoff, "Adaptive control by inverse th Dakota State University, Fargo, ND, USA, and modelling," 12 Asimolar conference on Circuits, Systems and Ph. D degree in Systems Science in 1971 from Computers, 1978. Michigan State University, East Lansing, MI, [14] B. Widrow, "Adaptive inverse control," in Proc. of Second IFAC USA. He was selected as a Fellow of IEEE in Workshop on Adaptive Systems in Control and Signal Processing, Lund January 2001 for his contributions to the Institute of technology, Lund, Sweden, pp:1-5, 1986. development and implementation of intelligent [15] B. Widrow and Gregory L. Plett, "Nonlinear adaptive inverse control," system techniques for power plants and power in Proc.of the 36th Conf. on Decision & Control. San Diego, California systems control. USA, pp. 1032-1037, 1997. He has been working in the area of power plants and power systems [16] B. Widrow and E. Walach, Adaptive Inverse Control. Prentice Hall PTR, control for over thirty years at Michigan State, Oregon State, University Upper Saddle River, NJ, 1996. of Houston, the Pennsylvania State University, and the Baylor University, [17] H. Chi-Li-Ma and K. Y. Lee, "Free-model based adaptive Inverse neuro- where he is Professor and Chairman of the Department of Electrical and controller for dynamic systems," Proceedings of the 37th IEEE Computer Engineering. Conference on Decision & Control, Tampa, Florida USA, 1998, pp. 507- His research interest includes control, operation, and planning of power 512. systems, computational intelligence, intelligent control and their [18] I. M. Galvan and J. M. Zaldivar, "Application of recurrent neural applications to power and energy systems, and modeling, simulation and networks in batch reactors, Part II: Nonlinear inverse and predictive control of renewable and distributed energy sources. control of the heat transfer fluid temperature," Chemical Engineering and Processing, 37:149-161, 1998. [19] F. M. Dias and A.M.Mota, "Direct inverse control of a kiln," 4th. Portuguese Conference on Automatic Control, Guimares, Portugal, 2000.