An optimization scheduling strategy for electric-heat-hydrogen integrated energy systems based on memory-enhanced deep reinforcement learning

High wind-solar penetration drives integrated energy systems (IESs) to act as cross-vector buffers that absorb surplus electricity via hybrid storage, or convert it into heat and hydrogen for cross-medium energy peak shaving. However, conventional mixed-integer linear programming (MILP) solvers and static incentive schemes often struggle with the high-dimensional, strongly coupled, non-convex, and time-varying nature of electric-heat-hydrogen scheduling. Thus, we propose an end-to-end optimization framework for a renewable electric-heat-hydrogen IES, and evaluate it through offline dispatch computations and cost settlement on a 24-h simulated case study. The proposed framework incorporates a coupled power–state of charge (SOC) dual penalty mechanism to ensure consistent storage operation over time. Additionally, it includes a bidirectional incentive-based demand response (B-IDR) mapping that is differentiable and can capture asymmetric feedback in response to price fluctuations. Furthermore, a long short-term memory (LSTM)-augmented maximum-entropy soft actor–critic (SAC) scheduler is utilized for stable and efficient control in this continuous and high-dimensional setting. Comparisons with other methods under identical settings show that the proposed method achieves lower operating costs and carbon emissions, as well as improved training stability.