An optimization control of thermal power combustion based on reinforcement learning
【摘要】：In this paper, a safe off-line training framework based on Deep Deterministic Policy is proposed to optimize the control of a continuous combustion process in real time. A simulator is constructed by a neural network with single input and double outputs,which to predict the expected technical indicators by the historical data. With such proposed prediction mode, a key step for its application is the transformation of the optimization problem into a Markov decision process with generalized information. The optimization framework underlying the deep deterministic policy gradient shows a great ability operating over continuous action spaces of high dimensions. Compared with the existing results, the proposed approach has the powerful generalization capacity in unexplored states. Finally, Numerical simulations are given to demonstrate the effectiveness of the proposed method.