A CNN-Transformer Model for Learning Electric Vehicle Routing Problem

作者：Zi-Long Liao, Jie Li, Junjie Fu · 发表于：Cybersecurity and Cyberforensics Conference · 年份：2025 · DOI：10.23919/ccc64809.2025.11178435

The electric vehicle routing problem (EVRP) poses significant challenges due to battery constraints, charging station dependencies, and complex route optimization requirements. Traditional heuristic methods struggle to balance computational efficiency with solution quality while existing deep learning models often lack robustness in handling large-scale or dynamically constrained scenarios. To address these limitations, we propose CNN - Transformer, a novel deep reinforcement learning model that integrates CNN, explicit sparse attention (ESA), and GRU. Our architecture employs a multi-head attention (MHA) mechanism with ESA to prioritize critical nodes and a GRU-enhanced decoder to capture long-term route dependencies. Comparative, ablation, and generalization experiments have demonstrated that our model outperforms existing models in terms of overall performance.