A physics-guided temporal convolutional learning with XGBoost integration for degradation-aware and robust electric vehicle range prediction

Accurate prediction of electric vehicle (EV) driving range is fundamental for alleviating range anxiety, improving route planning, and enabling reliable battery and energy management under real-world operating conditions. However, prevailing range prediction frameworks typically depend on instantaneous state-of-charge (SoC) and nominal battery capacity, implicitly assuming ideal battery behavior and neglecting the irreversible degradation induced by cycling, thermal stress, and heterogeneous charging practices. Conventional machine learning and deep learning methods often fail to represent these degradation dynamics, either because they ignore temporal dependencies or function as non-interpretable black-box models, thereby limiting their applicability in safety-critical battery management systems. This work introduces a Physics-Guided Temporal Convolutional Network integrated with XGBoost (PG-TCN-XGB)—a unified, degradation-aware predictive architecture designed to overcome these limitations. A key contribution of this study is the formulation of a Degradation Memory Index (DMI), a physics-derived cumulative descriptor that encodes long-term battery stress by jointly considering energy throughput, temperature deviation, and charging intensity. By embedding DMI directly into temporal feature learning, the model differentiates between identical SoC states arising from distinct degradation histories, thereby enabling physically consistent range estimates. The PG-TCN component leverages causal and dilated convolutions to model both short-horizon driving fluctuations and long-horizon behavioral and degradation patterns with high computational efficiency. These learned temporal embeddings are subsequently refined by an XGBoost regressor, which captures residual nonlinearities and mitigates noise sensitivity typical of real-world telemetry. Comprehensive evaluation using a large-scale real EV dataset containing battery characteristics, driving behavior, environmental variables, and charging-infrastructure indicators demonstrates that the proposed hybrid method significantly outperforms ensemble learning (Extra Trees, XGBoost) and deep learning baselines (LSTM, BiLSTM-Attention, CNN-LSTM). The PG-TCN-XGB model achieves superior performance with an RMSE of 0.037 and an R² of 0.978, with stability further confirmed through cross-validation. These results establish that combining physics-based degradation modeling with temporal deep learning and robust boosting strategies yields a substantial improvement in predictive accuracy, generalization across operating conditions, and interpretability. Overall, this study provides one of the first end-to-end physics-guided, degradation-aware temporal learning frameworks for EV range prediction, addressing critical gaps in current modeling approaches and offering a scalable, interpretable solution for next-generation EV energy management systems.