Eco-routing for electric logistic vehicles using deep reinforcement learning with graph attention networks

This paper proposes an eco-routing strategy for electric logistics vehicles (ELVs) with pickups and deliveries, leveraging deep reinforcement learning (DRL) enhanced with graph attention networks (GAT) to optimize routes for improving vehicle energy efficiency. First, the eco-routing problem for electric logistics vehicles (ELVs) with pickups and deliveries is formulated to account for vehicle dynamics, energy consumption, charging station distribution, and road slopes. Second, a DRL framework with GAT is proposed to efficiently solve the eco-routing problem. The policy network, built on the Transformer architecture, comprises an encoder, adjacency matrix, feature embedding, and decoder. The adjacency matrix represents the graph structure, capturing node relationships to enhance information propagation and minimize redundant computations. Finally, extensive simulations across diverse pickup and delivery scenarios for ELVs are conducted to validate the proposed method. The results show that the DRL-GAT method outperforms both existing learning-based and heuristic methods, achieving a 3%–9% improvement in routing optimality for the eco-routing problem, while exhibiting strong generalization capabilities. Moreover, the proposed eco-routing strategy significantly improves energy efficiency compared to the distance-minimizing baseline, achieving energy savings of up to 37%.