Multi-Agent Reinforcement Learning for Self-Healing Distribution Networks under Power Quality Constraints

Abstract

Distribution networks must restore healthy loads after faults without violating voltage, thermal, or radial topology constraints. This paper proposes an enhanced multi-agent reinforcement learning (MARL) framework for self-healing distribution networks under power quality constraints. The IEEE 33-bus radial distribution system is modeled in Python using pandapower, while the learning model is implemented in PyTorch. Fault isolation and service restoration are formulated as coordinated switching decisions over five normally open tie switches. A radiality-preserving repair mechanism is introduced to prevent looped post-restoration configurations, and a power-quality-aware reward function prioritizes restored load while penalizing voltage violations, line overloads, power losses, excessive switching, non-radiality, and power-flow failure. All feasible fault-restoration combinations are precomputed using pandapower and used as a cache during training. Across five fault scenarios, the enhanced MARL restored all recoverable loads, maintained radial topology, produced zero voltage violations and zero-line overloads, and achieved an average restored load of 82.48%, matching the exhaustive radial search benchmark. The average reward differed only slightly from exhaustive search (404.31 versus 404.53), while average losses were lower than the single-agent DQN baseline.