Real-Time Scheduling of High-Penetrated Renewable Power Systems: An Expert Knowledge and Reinforcement Learning Hybrid Approach
Sijun Du, Tao Ding, Yang Xiao, Jingyu Wan, Jun Liu, Fei Meng
Abstract
Modern power systems are undergoing a low-carbon and sustainable transition. The increasing penetration of renewable energy sources (RESs) poses significant challenges to the power system scheduling due to the associated uncertainties. Moreover, the integration of various flexible elements further complicates the scheduling problem. Therefore, rapid and accurate real-time scheduling methods are required to ensure the safe and stable operation of the power system. In this paper, a hybrid approach of expert knowledge and reinforcement learning (RL) is proposed to solve the real-time scheduling problem of the high-penetrated renewable power system. Firstly, a mathematical model for real-time scheduling of the high-penetrated renewable power system including flexible loads and energy storages (ESs) that integrates system operating costs and constraints, and RESs consumption is established and formulated as a Markov decision process. Subsequently, the proposed approach introduces expert knowledge as an intermediary between the power system and the RL agent, utilizing the optimized unit control sequence derived from the RL algorithm for scheduling decisions. Case studies conducted on the SG 126-bus system validate the effectiveness of the proposed approach and demonstrate its tremendous potential to facilitate RES consumption.