Forecasting electric vehicle charging demand using an explainable deep learning approach

The growing concern about climate change has made the reduction of carbon dioxide emissions a central objective in global sustainability strategies. In this context, the increasing diffusion of electric vehicles (EVs) introduces new challenges for energy systems. Although EVs reduce direct emissions from transportation, their charging demand increases electricity consumption and can indirectly contribute to CO_2 emissions depending on the electricity generation mix. This thesis investigates short-term electricity demand forecasting associated with EV charging, using the My Electric Avenue dataset. The raw data are cleaned, organized as time series, and transformed into a unified representation of electrical load. Several forecasting approaches are developed and compared: a calendar-based baseline, a feature-engineering approach with lag and rolling variables, and a sliding-window approach based on past observations. To respect the temporal nature of the data, chronological train-test splits and time-series cross-validation are adopted. Model performance is evaluated using RMSE, MAE, and R^2. Finally, SHAP and ALE are applied to interpret the trained models, while RExQUAL is used to quantitatively evaluate the quality of the explanations through association rules, Global Support, and Global Confidence. In this way, the thesis assesses both predictive accuracy and explanation reliability.