An Explainable and Uncertainty Aware Ensemble Framework for Electric Vehicle Charging Demand Forecasting

Abstract

The rapid expansion of electric vehicle (EV) adoption places increases pressure on power grids, making accurate charging demand forecasting a cornerstone of smartgrid. Yet most existing forecasting models look for prediction accuracy at the expense of transparency and uncertainty awareness two properties that operators and planners genuinely need when making real world infrastructure decisions. This paper presents Explain Charge, an ensemble forecasting framework that directly addresses these shortcomings. The framework combines XGBoost and Bidirectional LSTM in a dual branch architecture unified through a Ridge Regression meta learner, and introduces SHAP Guided Iterative Feature Refinement (SGIFR)a mechanism that actively uses SHAP importance scores during model development rather than treating explainability as an afterthought. Anomalous charging sessions are filtered upstream via Isolation Forest, and conformal prediction supplies statistically valid uncertainty intervals alongside every forecast. Evaluated on a multi station EV charging dataset, Explain Charge achieves a MAE of 0.4309 kWh and an $\mathbf{R}^{\mathbf{2}}$ of 0.9611, with 92% conformal coverage at the 90% nominal level, consistently outperforming standalone baselines. The results demonstrate that accuracy, interpretability and calibrated uncertainty are achievable together, offering a practical path toward trustworthy EV demand forecasting in operational smartgrid environments.