A new thermal power generation control in reinforcement learning

摘要

In this paper, a novel framework applying the reinforcement learning algorithms to thermal power generation control is proposed instead of traditional PID controllers. In order to implement a time series response by the neural network, we discrete the control model and embed it into the environment. By defining an episode as the adjustment process of the steady state to the setpoint, the agent can learn the whole embedding model instead of only accepting scalar error signals. The experiments have proven that proximal policy optimization framework shows the same robustness and fast response as the traditional PID control in the single input single output(SISO) system and performs better in the multiple input multiple output(MIMO) system.