Predicting Home EV Charging Practices Using Machine Learning

The rapid adoption of electric vehicles (EVs) demands advanced residential charging solutions, where user behavior varies widely due to factors like electricity tariffs and weather conditions. Such conditions make charging predictions particularly complex. This study addresses this challenge by predicting the time an EV remains connected to a residential charger using a dataset of 106,260 sessions. To this end, we propose the Connection Time Neural Model (CTNM), a deep neural network designed to model the complex dynamics of domestic charging, and we introduce the Weighted Error Metric (WEM), a novel metric that penalizes overestimations and under-estimations differently to reflect their real-world impacts on both grid management and user experience. Utilizing bidirectional charger data, we focus solely on connection time, bypassing energy prediction. CTNM is benchmarked against state-of-the-art methods (i.e., Random Forest, Dense Neural Network, XGBoost, and Support Vector Regression) using Mean Absolute Error, Root Mean Squared Error, and WEM as performance metrics. Results demonstrate CTNM’s superiority, reducing average error by 18%, and weighted error by 31% compared to Random Forest, thanks to its deep architecture and integration of contextual features, like variable tariffs and weather.