Safety-Guaranteed Energy Management in Networked Multienergy Microgrids: A Multi-Actor Single-Critic Deep Reinforcement Learning Approach

Distributed energy systems are shifting from stand-alone microgrids to networked multienergy microgrids (N-MEMGs), where energy management must handle multienergy coupling, renewable uncertainty, and strict device/network safety under distributed operation. A safety-guaranteed framework that couples a multi-actor single-critic multiagent deep reinforcement learning method with randomized ensembled double$Q$-learning (REDQ) and a model-based safety layer is proposed. Actors run locally at nodes for decentralized operation; a centralized REDQ-based critic module reduces$Q$-value overestimation and reliably guides policy updates; and the safety layer analytically enforces device, power balance, and multitimescale constraints, explicitly capturing inertia-induced effects along hydrogen device chains. On a five-node N-MEMG, the proposed method coordinates energy trading, hydrogen storage, and peak shaving, achieves zero observed safety violations over 12 000 evaluation steps, and outperforms mainstream baselines (multi-agent twin delayed deep deterministic policy gradient (MATD3), multiagent deep deterministic policy gradient, and multiagent soft actor critic) with an average daily adaptive normalized score of 98% across test days. The results point to a practical route to safe and scalable energy management in N-MEMGs.