Spatial–Temporal Deep Learning for Electric-Vehicle Charging Demand: An Exploratory Study of Graph Convolutional and LSTM Networks Performance

Abstract

Electric-vehicle (EV) charging is a localized, time-varying load that challenges distribution networks. This study offers practical insights into when spatial graph structure adds value beyond temporal context, utilizing real-world data and a transparent evaluation. We compare Long Short-Term Memory (LSTM) and Graph Convolutional Network (GCN) models for hourly EV-charging energy forecasting, based on 145,778 sessions recorded in Boulder, Colorado (2018–2023). After preprocessing and temporal alignment, temporal covariates (hour, day, month, year) and, when applicable, ZIP-code indicators were engineered. LSTMs were trained with 1 h and 24 h input windows, with or without ZIP features, and evaluated through 5-fold cross-validation. GCNs operated on hourly node–time tensors with a dense adjacency (no self-loops) and were trained on an 80/20 temporal split, using a 1% subsample for tractability. All models used Adam (lr = 0.005), early stopping, and ReduceLROnPlateau. Temporal context was the main driver of LSTM accuracy: 24 h inputs outperformed 1 h, while ZIP features improved only the shorter window. For GCNs, depth and node features shaped performance: a 2-layer GCN with ZIP features achieved the lowest RMSE (4.75 kWh), whereas a 6-layer GCN without node features reached the lowest MAE (2.46 kWh). Despite higher computational cost, GCNs captured spatial coupling effectively. Overall, 24 h LSTMs provide a strong and efficient baseline, while GCNs add value when spatial correlations are relevant. A hybrid GCN–LSTM architecture is a promising next step to jointly leverage spatial and temporal dependencies.