A Graph-based Deep Reinforcement Learning Framework for Autonomous Power Dispatch on Power Systems with Changing Topologies

Abstract

Increasing penetrations of renewable energy, flexible loads and distributed power supplies is prompting the mordern power system to be highly complex and uncertain. Besides, topology changes can greatly discount the effectiveness and real-time performance of traditional autonomous operation policies. Therefore, developing autonomous power dispatch methods is of great importance to the ensurance of modern power sytem economy and reliability. This paper proposes a novel graph-based deep reinforcement learning (DRL) framework for autonomous power dispatch considering topology changes. Based on the formulation of Markov decision process (MDP), a proximal policy optimization (PPO) algorithm with pre-training of imitation learning is adopted to obtain effective and timely power dispatch policies. Plus, to get the generalization ability to adopt to changing topologies caused by emergencies, maintenance plan and power grid construction, the GraphSAGE algorithm is embedded in the DRL agent to capture changing charcteristics of the power network. The case study is conducted on a modified IEEE 118-bus system and the results suggest good performance of the proposed framework.