Multi-Agent Reinforcement Learning for Scalable Multi-Objective Secure and Economic Operation of Flexible Distribution Networks With Congestion Mitigation

Urban distribution networks increasingly host heterogeneous operator-controlled distributed energy resources (DERs), storage, and remotely operable switches, demanding secure–economic operation that centralized mixed-integer formulations struggle to deliver in real time under uncertainty and scale. This study targets three gaps: unified coordination of a hybrid action space coupling continuous set-points with discrete topology reconfiguration; alignment of system-level objectives with local feasibility in multi-agent settings; and consistency between resource adjustments and switching under a training–execution decoupled paradigm. We propose a multi-agent deep reinforcement learning framework with a two-layer centralized-training– decentralized-execution (CTDE) architecture, where resource and topology agents share layer-wise critics and employ dual-headed policies; a global–local reward decomposition is embedded with a feasibility-aware action-projection operator that injects AC power-flow, voltage, and thermal constraints to guide exploration and ensure admissibility. On a modified IEEE 123-bus feeder, the approach outperforms a rolling Mixed-Integer Nonlinear Programming (MINLP) and Multi-Agent Reinforcement Learning (MARL) baselines, cutting total cost by 3.27%, losses by 23.69%, voltage-violation time by 23.44%, congestion rate by 29.17%, and load shedding by 52.94%, while reducing switching by 41.67% and achieving 0.05 s per-step inference (about 17× faster than a rolling MINLP baseline on the same 123-bus feeder on commodity hardware); under up to 20% forecast errors andN−1 contingencies it degrades gently and recovers faster.