A Cooperative Dispatch Algorithm for Hydrogen-Based Grid-Connection Microgrids: A Multi-Agent Reinforcement Learning Method

With the continuous development and progress of photovoltaic (PV) technology, the proportion of PV generation in microgrids has increased significantly, making microgrids more low-carbon and environmentally friendly. However, the high PV penetration can lead to excessive fluctuations in the exchange power at the point of common coupling (PCC), resulting in transmission line overheating and posing risks to the microgrid's stability. Aiming to address the active power fluctuations at PCC caused by high PV generation, the paper proposes a new cooperative dispatch algorithm based on multi-agent reinforcement learning. Our algorithm improves the training approach of dueling double deep Q-network by splitting the original reward function into a new auxiliary network. The auxiliary network guides the update direction of the main network during the training process, enabling the power system to mitigate power fluctuations while satisfying other physical constraints. Agents collaboratively learn optimal power dispatch strategies for hydrogen storage systems via localized communication with neighboring agents. We compare our proposed algorithm with baseline algorithms, and simulation results demonstrate its superior performance in mitigating active power fluctuations at PCC and maintaining stable load status for hydrogen energy storage systems.