Reinforcement Learning with Model Predictive Control for Coordinated Operation of Power Plant

To address the challenges of rapid and accurate regulation of power generation load due to significant model nonlinearity, strong inertia, and substantial variable coupling in the load control process of boiler and turbine units in coal-fired power plants, this paper proposes a control policy that combines model predictive control with reinforcement learning based on a multi-operating condition model. By making real-time adjustments according to the operating state, the policy corrects model predictive control towards optimal control, thereby achieving optimal variable load control of a coal-fired power unit. A case study of a typical 300MW power plant in China demonstrates the superior control performance of the proposed method compared to traditional predictive control models.

1. Introduction

With the rapid advancement of new energy generation technologies, China's electric power structure is continuously shifting towards cleaner power generation. Although the proportion of thermal power in China's total energy capacity has declined, coal-fired power will continue to dominate for the foreseeable future. Boiler-turbine units, essential components of coal- fired power plants, must meet power load demands while maintaining stable operating parameters. However, challenges arise due to strong variable coupling, significant model nonlinearity, and complex control input constraints. Various control methods have been employed to tackle these issues, such as Proportional Integral Derivative (PID) control [1], Model Predictive Control (MPC) [2], Active Disturbance Rejection Control (ADRC) [3], and fuzzy control [4]. Despite their effectiveness, these methods are computationally intensive when handling high-dimensional control processes, underscoring the urgent need for an efficient control scheme capable of managing complex optimal control problems.