Reinforcement Learning-based Optimization for Cold Chain Electric Vehicle Routing

To address the range limitations and hard time window constraints in cold chain electric vehicle delivery, this article proposes EGT-TA, an end-to-end deep reinforcement learning framework designed to minimize comprehensive energy consumption costs while satisfying complex constraints. The model utilizes an Edge-augmented Graph Transformer (EGT) encoder to deeply integrate node and topological features. Furthermore, a dynamic urgency-biased mechanism is incorporated into the decoder to dynamically adjust node selection probabilities by perceiving remaining time windows, significantly enhancing the model's control over timeliness. Additionally, a constrained POMO training algorithm based on Lagrangian relaxation is introduced to guide the agent in adaptively balancing solution feasibility and optimality during exploration. Experimental results demonstrate that the proposed model outperforms Genetic Algorithms and classic Attention Models in solution quality across instances of varying scales.