Multi-Microgrid Energy Management Strategy Based on Multi-Agent Deep Reinforcement Learning with Prioritized Experience Replay

The multi-microgrid (MMG) system has attracted more and more attention due to its low carbon emissions and flexibility. This paper proposes a multi-agent reinforcement learning algorithm for real-time energy management of an MMG. In this problem, the MMG is connected to a distribution network (DN). The distribution network operator (DSO) and each microgrid (MG) are modeled as autonomous agents. Each agent makes decisions to suit its interests based on local information. The decision-making problem of multiple agents is modeled as a Markov game and solved by the prioritized multi-agent deep deterministic policy gradient (PMADDPG), where only local observation is required for each agent to make decisions, the centralized training mechanism is applied to learn coordination strategy, and a prioritized experience replay (PER) strategy is adopted to improve learning efficiency. The proposed method can deal with the non-stationary problems in the process of a multi-agent game with partial observable information. In the execution stage, all trained agents are deployed in a distributed manner and make decisions in real time. Simulation results show that according to the proposed method, the training process of a multi-agent game is accelerated, and multiple agents can make optimal decisions only by local information.

Automated summary

This paper proposes a multi-agent reinforcement learning algorithm for real-time energy management of a multi-microgrid (MMG) system. The MMG is connected to a distribution network, and the distribution network operator and each microgrid are modeled as autonomous agents. Each agent makes decisions based on local information, and the problem is modeled as a Markov game and solved using the prioritized multi-agent deep deterministic policy gradient (PMADDPG) algorithm.