Physics-augmented safe reinforcement learning for overload mitigation in distribution networks under weather-sensitive thermal constraints

The widespread integration of distributed PVs is increasingly constrained by line overloads, emerging from reverse power flows and heterogeneous loads. These issues are especially pronounced in high-R/X distribution networks with weather-sensitive conductor thermal dynamics. Yet traditional static paradigms fail to leverage inherent dynamic headroom and cost-effective flexibility. To overcome these limitations, this paper proposes a dynamic thermal-aware distribution flexibility management (DFM) framework that explicitly integrates weather-sensitive line thermal dynamics with physics-augmented and safety-aware demand-side control. First, we develop a dynamic-thermal-constrained power flow model (DT-PF) that merges dynamic line rating (DLR) with time-varying ambient conditions, yielding physics-consistent and real-time estimation of branch ampacity. Built upon this model, the DFM orchestrates aggregated heating, ventilation and air conditioning (HVAC) loads and battery energy-storage systems (BESS) for economic operation. Subsequently, a hybrid safe-reinforcement-learning (SRL) approach is proposed to solve the high-dimensional and non-convex problem under ambient uncertainty. On the one hand, the SRL agent employs the action shielding mechanism for instantaneous enforcement of device-level constraints. On the other hand, it incorporates the physics-guided safety-critics to guide the agent satisfying network-wide thermal safety. Simulation studies based on practical data in a city in southern China effectively alleviate overload incidents, simultaneously upholding user-comfort thresholds and network thermal security under uncertainty.