A Multi-Source and Data-Driven Deep Learning Framework for Forecasting EV Charging Demand

Urban mobility is experiencing a major shift with the rising popularity of electric vehicles (EVs). To meet the growing demand for EVs, charging stations must offer sufficient coverage, therefore analyzing the current infrastructure is key to be able to forecast future demand. This study presents an innovative data-driven deep learning framework to predict charging demand, integrating data from public parking lot occupancy, charging station usage, and crowdsourced point of interest (POI) popularity. Our paper provides two significant contributions: first, we study the correlation between three data types - charging and parking occupancy, and crowdsourced data. Secondly, we propose a predictive model for EV charging demand integrating parking occupancy and crowdsourced data, disconnecting the prediction from the historical data of the selected charging stations. Consequently, this model is wellsuited for application where there are areas lacking charging infrastructure, with the objective of installing new charging stations without the prior knowledge of the demand. Sequence-to-Sequence Recurrent Neural Networks (seq2seq RNNs) were applied with various learning time windows to predict the next 24-hour charging usage pattern. The model, trained on a zonelevel and used a 72-hour data window, demonstrated better performance compared to other models, achieving a Root Mean Squared Error (RMSE) of 6.75, a Mean Absolute Error (MAE) of 5.02, and an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score of 0.80 on the test data.