ATLAS-EV: An Adaptive Meta-learning Framework for Electric Vehicle Routing Problem with Time Windows

The rapid adoption of electric vehicles (EVs) in commercial fleets necessitates sophisticated routing solutions for the Electric Vehicle Routing Problem with Time Windows (EVRPTW). Existing approaches lack adaptability across diverse problem instances, creating demand for integrated frameworks. This research introduces ATLAS-EV (Adaptive Training and Learning Algorithm Selection for Electric Vehicles), a meta-learning framework implementing three-level adaptation: (1) algorithm selection adaptation through supervised learning mapping problem features to optimal meta-heuristics, (2) operator probability adaptation using reinforcement learning where successful operators receive increased selection probability, and (3) parameter adaptation automatically calibrating algorithm-specific parameters based on problem characteristics. The framework incorporates comprehensive feature engineering with seven basic features and six landmarking features, alongside three specialized meta-heuristics: HR-GA for complex route structures, CACO for charging-critical scenarios, and AE-PSO for energy-intensive routes. SMOTE addresses class imbalance, achieving balanced 33.33\% representation across algorithms. Validation on Schneider benchmark instances demonstrates meta-learning classification accuracy of 87.5\% with Random Forest and perfect ROC AUC for algorithm selection. Feature importance analysis reveals normalized demand and time window characteristics as primary selection drivers. A hypothetical case study projects 20.5\% improvement in on-time delivery and 22.9\% energy cost reduction. ATLAS-EV advances electric vehicle routing optimization by providing an adaptive framework that eliminates manual parameter tuning while effectively solving diverse EVRPTW instances.