Expert-guided attention-MAPPO with CBF-based safety correction for low-carbon economic dispatch of hydrogen-integrated energy systems

To address multi-energy coupling, uncertain renewable generation, and safety-constrained operation in hydrogen-integrated energy systems, this paper proposes an expert-guided safe multi-agent reinforcement-learning framework for low-carbon economic dispatch. The system is decomposed into three coordinated agents representing the electrical, heating/cooling, and hydrogen subsystems. A physics-based action decomposition distinguishes active control variables from passive energy-balance variables, thereby reducing the effective exploration space. The multi-agent proximal policy optimization-based scheduling framework incorporates behavior-cloning pretraining with expert trajectories and an annealing mechanism that gradually reduces expert supervision during training. An attention-based representation module is included to model dependencies among heterogeneous energy states. A discrete-time control barrier function safety layer projects nominal actions onto the feasible set before execution and provides hard guarantees for the constraints explicitly included in the control barrier function (CBF) formulation. Across five independent training seeds and 20 fixed test days, the complete method achieved an average operating cost of 9846.32±326.88 Chinese Yuan (CNY)/day, representing a 17.44% reduction relative to proximal policy optimization and a 14.68% gap relative to the full-information mixed integer linear programming benchmark. The CBF safety layer eliminated all violations of the constraints explicitly covered by its formulation. Component-wise ablation results show that CBF provides the largest reduction in cost and constraint violations, whereas the attention module alone does not reduce operating cost under the current configuration. Tests under 80% capacity, 120% load, and doubled disturbance standard deviations further evaluate robustness within the examined operating range.