(108d) Comparison of State-of-the-Art Dynamic Machine Learning Methods for MPC of Coal-Fired Utility Generator Performance

The identity of the world’s electricity generating sources is evolving at an unprecedented rate. Massive additions of variable renewable energy sources (VREs) are fundamentally changing operating methods of existing fossil-fuel power plants, as they attempt to maintain reliability and meet demand [1], [2]. Many large coal-fired generators originally designed as baseloaded units are now performing load following in a manner never anticipated at ramp rates once thought to be impossible [3], [4]. This significant departure from design conditions can lead to substandard operating performance, decreased efficiency, and increased emission rates [5]. Efforts by many countries to significantly reduce their carbon footprints further escalates the need to reduce the heat rate of thermal utility generators in order to achieve these goals. As these practices are adopted by more generation facilities, advanced control strategies must be developed to maintain appropriate operation. Real-time artificial intelligence optimization systems have been shown to consistently reduce emission rates of coal-fired power plants in closed-loop control and are a preferred method for reducing heat rate. [6]–[8], but observed benefits are limited during transient operation. Model predictive control (MPC) is a proven method of optimizing dynamic systems during transient operation [9], [10]. To date, the application of MPC to combustion control has been precluded by the complexity of dynamically modeling the combustion process. Advancements in data-driven machine learning modeling methods are reducing the barrier to entry of model generation and application of MPC to combustion control. This work examines current state-of-the-art dynamic machine learning modeling methods and their suitability for use within an MPC framework to minimize heat rate and NOx emission rates at a coal-fired utility generator. Existing dynamic modeling methods, including forms of autoregressive models (VAR, ARX, NARX), recurrent neural networks (RNN, LSTM, bidirectional RNN and LSTM, CTRNN, DBN), and recurrent support vector machines, are compared based on prediction accuracy and other criteria over multiple time horizons on operational data from a 500 MW coal-fired utility generator. Advantages and disadvantages of each method are discussed. Appropriate modeling methods for use in future work of developing and operating an MPC system in closed-loop control at this utility generator are identified.

[1] W. Cole et al., “Variable Renewable Energy in Long-Term Planning Models : A Multi-Model Perspective Variable Renewable Energy in Long-term Planning Models : A Multi-model Perspective,” Golden, 2017.

[2] P. Denholm, G. Brinkman, and T. Mai, “How low can you go? The importance of quantifying minimum generation levels for renewable integration,” Energy Policy, vol. 115, no. October 2017, pp. 249–257, 2018.

[3] E. Moiseeva, S. Wogrin, and M. R. Hesamzadeh, “Generation flexibility in ramp rates: Strategic behavior and lessons for electricity market design,” Eur. J. Oper. Res., vol. 261, no. 2, pp. 755–771, 2017.

[4] S. M. Safdarnejad, J. F. Tuttle, and K. M. Powell, “Dynamic modeling and optimization of a coal-fired utility boiler to forecast and minimize NOx and CO emissions simultaneously,” Comput. Chem. Eng., vol. 124, pp. 62–79, 2019.

[5] J. F. Tuttle and K. M. Powell, “Analysis of a thermal generator’s participation in the Western Energy Imbalance Market and the resulting effects on overall performance and emissions,” Electr. J., vol. 32, no. 5, pp. 38–46, 2019.

[6] J. F. Tuttle, R. Vesel, S. Alagarsamy, L. D. Blackburn, and K. Powell, “Sustainable NOx emission reduction at a coal-fired power station through the use of online neural network modeling and particle swarm optimization,” Control Eng. Pract., vol. 93, no. September, 2019.

[7] R. Cass and B. Radl, “ADAPTIVE PROCESS OPTIMIZATION USING FUNCTIONAL-LINK NETWORKS AND EVOLUTIONARY OPTIMIZATION,” Control Eng. Pract., vol. 4, no. 11, pp. 1579–1584, 1996.

[8] D. Hasler and W. Rosenquist, “Coal-Fired Power Plant Heat Rate Reductions,” 2009.

[9] J. D. Hedengren, R. A. Shishavan, K. M. Powell, and T. F. Edgar, “Nonlinear modeling, estimation and predictive control in APMonitor,” Comput. Chem. Eng., vol. 70, pp. 133–148, 2014.

[10] J. B. Rawlings, D. Q. Mayne, and M. M. Diehl, Model Predictive Control: Theory, Computation, and Design, 2nd ed. Madison: Nob Hill Publishing, 2017.