Uncertainty-Aware Machine Learning for Electric Vehicle Load Forecasting

The integration of electric vehicles (EVs) into power grids introduces significant uncertainty in energy demand forecasting due to unpredictable charging behaviors, variable driving patterns, and temporal dependencies. This uncertainty, if inadequately addressed, can lead to grid instability, inefficient resource allocation, and operational failures. This paper presents a comprehensive uncertainty-aware benchmarking framework that systematically evaluates six machine learning (ML) models namely, Ensemble, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Feedforward Neural Network (FNN), Bidirectional Long Short-Term Memory (BiLSTM), and Gated Recurrent Unit (GRU) under controlled experimental conditions with explicitly modeled EV load uncertainty. The uncertainty is formulated as both parametric and non-parametric components, integrating them into a unified forecasting optimization framework. Using real-world Institute of Electrical and Electronics Engineers (IEEE) datasets adapted for Indian EV load patterns, all models are trained and tested under identical time-series split conditions of 480 hourly samples with standardized hyperparameter tuning protocols. Advanced multidimensional visualization techniques including three-dimensional (3D) and five-dimensional (5D) scatter analysis reveal model sensitivity to uncertainty across temperature, session duration, and temporal factors. Results demonstrate that XGBoost achieves superior performance with Root Mean Square Error (RMSE) of 0.924 and accuracy of 85.3%, showing robust generalization under uncertainty, followed closely by SVM with RMSE of 0.917 and accuracy of 85.6%. Deep learning models exhibit higher computational costs but show promise with adequate training resources. This work provides actionable insights for deploying artificial-intelligence (AI)-driven forecasting systems in smart grid environments under realistic uncertainty constraints.