Data-Driven Load Profile Forecasting for EV Charging Stations Leveraging Spatial Dependency Modeling

Abstract

Accurate forecasting of energy demand at electric vehicle charging stations is critical for efficient energy management, reducing grid stress, and supporting the expansion of sustainable transport infrastructure. In this study, we introduce a novel forecasting approach that leverages spatial dependencies across a network of charging stations to improve the predictions of energy demand. Using EMOTION data from the city of Galantina, Italy, our model integrates information from nearby stations, capturing local demand patterns and spatial correlations through the use of two advanced recurrent neural network models, Long Short-Term Memory network (LSTM) and Graph Convolutional Long Short Term Memory (GCLSTM). The model is designed to predict energy demand by learning dynamic interdependencies within the city's charging network, where each station's demand is influenced by spatially adjacent stations and its distance to the city center. The LSTM model, which treats charging stations as independent entities, achieved an overall Mean Absolute Error (MAE) of 0.34, while GCLSTM reached 0.45 MAE. However, GCLSTM outperformed LSTM for stations located near other charging stations, demonstrating its ability to take advantage of spatial dependencies. This suggests that geobased prediction models are particularly beneficial for dense urban areas with multiple charging stations, such as city centers.