Spatio-Temporal Graph Convolutional Networks for EV Charging Demand Forecasting Using Real-World Multi-Modal Data Integration

TL;DR

This work introduces TW-GCN, a spatio-temporal graph convolutional network framework designed to forecast EV charging demand by fusing real-world traffic, weather, and POI data with ChargePoint station logs in Tennessee. The model constructs geospatial and DTW-based adjacency matrices, uses temporal stacking of geo- and demand-related signals, and applies region-aware clustering with MLPs to generate station-level energy forecasts. Through extensive experiments across lag hours, cluster counts, and temporal windows, TW-GCN demonstrates superior robustness and accuracy relative to baselines, with mid-horizon forecasts (approximately 3 hours) offering the best trade-off between responsiveness and stability, and certain regions (e.g., East Tennessee) yielding particularly reliable predictions. The findings support data-driven decisions for EV infrastructure deployment and grid management, highlighting how traffic, weather, and POIs shape charging patterns and how phased deployment can strategically grow a resilient charging network.

Abstract

Transportation remains a major contributor to greenhouse gas emissions, highlighting the urgency of transitioning toward sustainable alternatives such as electric vehicles (EVs). Yet, uneven spatial distribution and irregular utilization of charging infrastructure create challenges for both power grid stability and investment planning. This study introduces TW-GCN, a spatio-temporal forecasting framework that combines Graph Convolutional Networks with temporal architectures to predict EV charging demand in Tennessee, United States (U.S.). We utilize real-world traffic flows, weather conditions, and proprietary data provided by one of the largest EV infrastructure company in the U.S. to capture both spatial dependencies and temporal dynamics. Extensive experiments across varying lag horizons, clustering strategies, and sequence lengths reveal that mid-horizon (3-hour) forecasts achieve the best balance between responsiveness and stability, with 1DCNN consistently outperforming other temporal models. Regional analysis shows disparities in predictive accuracy across East, Middle, and West Tennessee, reflecting how station density, population, and local demand variability shape model performance. The proposed TW-GCN framework advances the integration of data-driven intelligence into EV infrastructure planning, supporting both sustainable mobility transitions and resilient grid management.

Figures (7)

• Figure 1: The TW-GCN architecture processes spatiotemporal data (B, T, N, F) with added contextual features (weather, traffic, POIs). It employs GCN layers for spatial modeling at each timestep, transforming node features from (B$\times$N, F) to (B$\times$N, H), using Dynamic time warping (DTW) for temporal similarity and Haversine distance to transform arrays into adjacency matrices: Geographic (x_geo) and demographic (x_dem) features are later fused via a weighted sum ($\alpha$). The stacked outputs (B, T, N, H) feed into a temporal model. K-Means clustering on the final timestep's features identifies distinct regions, and K-MLPs generate region-specific predictions (B, T, N, 1), resulting in robust, region-aware forecasting.
• Figure 2: Distribution of error metrics (MAE, RMSE, MSE, SMAPE) across tuned deep learning models. Each boxplot is displayed on a logarithmic scale to highlight performance differences, with lower values indicating better accuracy.
• Figure 3: Comparison of TW-GCN models against baseline machine learning models across error metrics (MAE, RMSE, MSE, SMAPE). TW-GCN models are shown as boxplots, while baseline ML models are shown as overlaid stripplots for direct comparison.
• Figure 4: Geographic distribution and energy consumption of EV charging stations for January 2024. The map visualizes 61 unique stations with 2,522 recorded charging events, highlighting spatial patterns of energy usage. Average energy per record is 6.10 kWh, and total energy consumption across all stations is 15,381 kWh. Weather conditions during this period are included for context, though correlations with energy usage are low, suggesting that EV charging demand is primarily driven by user behavior rather than weather factors.
• Figure 5: Focused analysis of East Tennessee with best-performing station green star. Average stations are shown in yellow. Principal cities are labeled with blue triangles.