A scalable graph reinforcement learning algorithm based stochastic dynamic dispatch of power system under high penetration of renewable energy

Abstract

Due to the increasing penetration of renewable energy source, power system is facing significant uncertainties. How to fully consider those uncertainties in dynamic economic dispatch (DED) has become a crucial problem to the safe and economic operation of power system. Reinforcement learning (RL) based approaches can provide the dispatch policy in response to uncertainty. However, current RL uses traditional Euclidean data representation, which greatly reduces the scalability and computational efficiency of economic dispatch algorithms. To address such obstacle, this paper develops a novel graph reinforcement learning (GRL) method for DED of power system. Firstly, DED is formulated as a Markov decision process and formulated as a dynamic sequential decision problem. Secondly, a novel graph-based representation of system state is proposed. Using graph data to represent dispatch operation data with non-Euclidean characteristics can effectively capture the implicit correlation of the uncertainty corresponding to the system topology. Thirdly, a GRL algorithm is proposed to learn the optimal policy which maps graph represented system state to DED decision. Compared with the traditional deep reinforcement learning (DRL), the GRL proposed in this paper has certain generalization ability and scalability, and can achieve higher quality solutions in online operation. Case studies illustrate that the optimality of the proposed method is 98.04%, which is 15.13% higher than that of the existing learning methods. The algorithm is scalable and improves the sample efficiency.