Process Knowledge-Guided Optimization Control for Once-Through Boiler-Turbine Units Based on Multi-Agent Reinforcement Learning

Abstract

For the coal-fired power unit, traditional centralized optimization control frameworks face challenges in achieving fast load response due to heavy computation and long online calculation times, especially for devices with limited computing power. This paper proposes a process knowledge-guided distributed optimization control framework for once-through boiler-turbine unit using multiagent deep reinforcement learning. In this framework, a centralized training distributed execution multiagent deep reinforcement learning algorithm is employed to obtain the optimal controllers of the once-through boiler-turbine unit, by dividing the coordinated control system into three subsystems and modeling the control process as a fully cooperative multiagent Markov decision process. Moreover, the process knowledge represented by a low-precision process model is used to guide and improve the training of multiagent deep reinforcement learning by distributed model predictive control algorithm generating the initial control actions and the designed action fusion strategy. Finally, the effectiveness of the process knowledge-guided optimization control framework is verified by the simulation platform, and the results show that the proposed algorithm has a faster speed and better control effect than the compared algorithms.