Safe Deep Reinforcement Learning for Hydrogen-Blended Integrated Electricity-Gas Systems with Wind Power and Stochastic EV Load

This study develops a safety-aware deep reinforcement learning (DRL) scheme for online operational decision-making in hydrogen-blended integrated electricity-gas systems (HB-IEGS) with wind power and stochastic electric vehicle (EV) loads. The problem is converted to a constrained Markov decision process (CMDP), and a constrained soft actor-critic (C-SAC) algorithm is adopted within a physics-guided environment that captures power flow, gas dynamics, hydrogen blending, and storage constraints. Numerical results on a 14-bus/8-node system show that the proposed method achieves high feasibility and significantly reduces online computation time compared with optimization-based benchmarks. Moreover, a moderate hydrogen blending ratio is found to provide the best trade-off between operating cost and carbon emissions, with 20% blending yielding the most favorable performance.