Session and energy forecasting for electric vehicle charging station using custom weighted ensemble machine learning
R. Nagha Akshayaa, S. A. Sajidha, R. Radha
Abstract
• The proposed ensemble learning framework effectively predicts total energy delivered and session duration in electric vehicle (EV) charging station. • Implemented diverse machine learning algorithms and employed a novel weighted strategy. • These results have significant implications for EV charging infrastructure. • Leads to better energy allocation, optimized charging schedules, and increased grid efficiency. • Contributes to a more sustainable, data-driven EV ecosystem and reducing grid stress. The rapid adoption of electric vehicles (EVs) presents a significant challenge for EV charging stations in efficiently managing and allocating energy. This study develops an advanced ensemble learning framework to predict total energy delivered and session duration, leveraging charging session data and vehicle attributes from a U.S. Department of Energy and CALSTART dataset. Data from Electric Vehicle Supply Equipment (EVSE), onboard systems, and utility submeters enable precise vehicle mapping via Vehicle ID. A custom-weighted ensemble model, combining Random Forest, CatBoost, LightBGM, SVR, XGBoost, and neural networks, employing a novel weight assignment strategy has been proposed. The result of this research achieves a state-of-the-art R 2 score of 99.57 % and a SMAPE of 5.88 % for total energy delivered, as well as an R 2 score of 82.2 % and a SMAPE of 9.81 % for session duration—significantly outperforming existing models. The attributes contributing to the prediction are also studied using Shapley Additive exPlanations (SHAP). Hence, the operations of electric vehicle charging stations can be optimised with the proposed model for real-time scenarios .