Adaptive transformer disconnection control under disaster-induced uncertainty

Extreme natural disasters introduce rapid load fluctuations, partial sensing failures, and communication degradation that fundamentally challenge the reliability of transformer protection systems. This paper proposes PODAR, a partial-observation deep adaptive relaying framework that formulates shutdown control as a cooperative POMDP and learns robust transformer protection strategies through multiagent reinforcement learning. PODAR integrates a multimodal observation encoder, a belief-state reconstruction module, a safety-constrained action projection mechanism, and a Nash-stable coordination layer that mitigates inconsistent or conflicting agent decisions. Extensive disaster-driven experiments demonstrate that PODAR maintains high shutdown accuracy under severe observation loss, significantly reduces false and missed trips compared with state-of-the-art learning and rule-based baselines, and preserves system-level selectivity through coordinated action refinement. The results highlight the feasibility of embedding learning-driven algorithms into protective relaying to achieve adaptive, interpretable, and safety-guaranteed operation under extreme uncertainty.