Accurate short-term forecasts of the power drawn by individual EVs charging sessions are critical for sizing and operating public charging hubs, yet hub managers seldom know battery capacity, initial State Of Charge (SOC) or charging-algorithm details for the vehicles they serve. This paper explores how much predictive skill can be extracted when the only information available is the historical power-time profile of past sessions. A real-world data set of 719 209 charging events is cleaned, filtered and analysed. Several neural time-series architectures are trained to predict the second half of a recharge event from its first half, with the best performer scoring 41 Symmetric Mean Absolute Percentage Error (SMAPE). Scaling the training set or deliberately over-fitting on single-user subsets fails to improve accuracy, indicating that the information missing from the operator’s point of view fundamentally constrains forecastability. The results therefore motivate the integration of lightweight EV metadata (e.g., battery size or state-of-charge) or external contextual signals to unlock higher-fidelity power-demand forecasting.
Predicting EV Recharges from the Point of View of Charging Hub Managers
Rios, Federico;Colombo, Cristian Giovanni;Longo, Michela;Zaninelli, Dario;Leva, Sonia
2025-01-01
Abstract
Accurate short-term forecasts of the power drawn by individual EVs charging sessions are critical for sizing and operating public charging hubs, yet hub managers seldom know battery capacity, initial State Of Charge (SOC) or charging-algorithm details for the vehicles they serve. This paper explores how much predictive skill can be extracted when the only information available is the historical power-time profile of past sessions. A real-world data set of 719 209 charging events is cleaned, filtered and analysed. Several neural time-series architectures are trained to predict the second half of a recharge event from its first half, with the best performer scoring 41 Symmetric Mean Absolute Percentage Error (SMAPE). Scaling the training set or deliberately over-fitting on single-user subsets fails to improve accuracy, indicating that the information missing from the operator’s point of view fundamentally constrains forecastability. The results therefore motivate the integration of lightweight EV metadata (e.g., battery size or state-of-charge) or external contextual signals to unlock higher-fidelity power-demand forecasting.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
