Prediction-aided hierarchical and safe reinforcement learning for cold-chain electric vehicle routing

Cold-chain electric vehicle routing is a challenging problem in sustainable logistics because routing decisions must jointly consider traffic congestion, battery depletion, charging availability, refrigeration demand, thermal safety, and customer time windows. These tightly coupled factors make conventional static routing models inadequate, particularly in dynamic refrigerated distribution environments where future operating conditions and service risks must be anticipated. This paper proposes a Prediction-aided Hierarchical and Safe Reinforcement Learning (PHASE) framework for cold-chain electric vehicle routing. The proposed framework integrates prediction-aided state representation, hierarchical decision-making, and safety-oriented policy learning into a unified routing paradigm. By embedding future traffic, charging, and thermal information into the decision process, PHASE enables refrigerated electric vehicles to make forward-looking operational choices. The hierarchical structure distinguishes among service, charging, return, and protection behaviors, while the safety mechanism reduces infeasible actions caused by low state-of-charge, excessive cargo-temperature deviation, or severe service delay. The proposed framework improves the adaptability and robustness of routing policies under time-varying operating conditions and supports coordinated optimization of punctuality, energy efficiency, charging efficiency, and cold-chain safety. PHASE is particularly suitable for large-scale refrigerated logistics systems in which refrigeration energy consumption and spoilage risk must be explicitly incorporated into routing intelligence. The framework provides a scalable and practical solution for intelligent cold-chain electric distribution under complex dynamic constraints.