Real-time energy management strategy of the hydrogen-coupled microgrid based on Informer model prediction results and deep reinforcement learning

Abstract

Abstract Microgrids (MG) powered by emerging distributed energy resources (DERs) are increasingly pivotal in driving the decarbonization of the power system. Due to the uncertainty of renewable energy sources, electricity generated from wind and solar is typically stored, either in batteries or in the form of high-pressure hydrogen gas within hydrogen storage tanks to meet the refuelling requirements of fuel cell vehicles (FCVs). For FCVs, the substantial operational expenses associated with hydrogen refuelling stations pose a significant barrier to their widespread adoption, with a notable portion of these costs attributed to hydrogen transportation. To address this challenge, this paper employs the Informer model based on a transformer for day-ahead prediction of photovoltaic power generation, as well as load demand forecasting. Subsequently, the Twin-delayed deep deterministic (TD3) policy gradient algorithm is used for real-time energy management of the microgrid. While ensuring that the State of Charge (SOC) remains above 45% daily to meet future hydrogen demands, the MG pursues an optimal strategy to minimize the total daily cost, including electricity procurement, carbon emissions and equipment degradation. Compared to outcomes derived from rule-based approaches, the algorithm introduced in this paper minimizes solar waste and equipment degradation, facilitating hydrogen production during low-cost electricity periods and enhancing economic feasibility.