Safe Multi-Agent Deep Reinforcement Learning-Based Energy Management for Multi-Microgrid Energy Communities

Abstract

Multi-microgrid (MMG) systems enhance efficient integration of renewables and emerging loads into modern sustainable communities. By facilitating energy sharing among MGs, MMG systems can enhance reliability, improve efficiency, and promote economic operation. In this paper, a projection based safe multi-agent deep reinforcement learning (DRL) algorithm is developed for the decentralized energy management of multi-microgrid energy communities. The safe projection layer integrated into DRL training effectively restricts the agent’s actions to the feasible space, thereby preventing any violations of physical constraints during system operation. The decentralized application of the proposed safe DRL methods achieves the optimal energy management without high-fidelity communication, thereby facilitating practical deployment. The modified MMG test systems are adopted to test the proposed method, and the testing results could verify the effectiveness of the proposed energy management system in achieving cost-effective operation while guaranteeing 100% safety and eliminating the dependence on high-fidelity communication.