Deployment-oriented evaluation of temporal and spatial forecasting approaches for electric vehicle charging demand in energy systems

Document Type

Article

Source of Publication

Energy Conversion and Management X

Publication Date

9-1-2026

Abstract

Accurate short-term electric vehicle (EV) charging demand forecasting is important for charging infrastructure operation, grid management, and energy-system planning. This study presents a deployment-oriented and reproducible evaluation of temporal, spatial, and unified spatio-temporal forecasting approaches for day-ahead EV charging demand prediction. Using publicly available charging-session data aggregated at hourly resolution across ZIP-code regions, we compare persistence and ARIMA baselines, XGBoost, Long Short-Term Memory (LSTM) networks, Graph Convolutional Networks (GCNs), and a unified GCN+LSTM architecture under a consistent preprocessing pipeline, leakage-free validation protocol, and rolling-origin evaluation framework. For the Boulder ZIP-code dataset considered in this study, temporal information provided the dominant predictive signal. Under rolling-origin evaluation, XGBoost achieved the lowest RMSE (14.612 kWh), with GCN+LSTM (14.725 kWh) and LSTM (14.841 kWh) achieving comparable performance, while the spatial-only GCN achieved the lowest MAE (8.542 kWh) but exhibited greater sensitivity to large forecast deviations. Calendar-derived temporal covariates consistently improved forecasting accuracy across model classes, whereas extending the historical input window beyond 24 h provided limited additional benefit. Although spatial modeling did not consistently improve average forecasting accuracy, the unified GCN+LSTM model exhibited slightly lower RMSE variability and marginally lower high-quantile forecasting errors under the evaluated conditions. Overall, the results highlight the importance of reproducible preprocessing, deployment-oriented evaluation, and controlled benchmarking when assessing forecasting methods for EV charging infrastructure and energy-system applications.

ISSN

2590-1745

Publisher

Elsevier BV

Volume

31

Disciplines

Electrical and Computer Engineering

Keywords

Day-ahead forecasting, Electric vehicle charging demand, Energy demand forecasting, Energy systems modeling, Reproducible evaluation, Spatial–temporal analysis

Scopus ID

105043366777

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Indexed in Scopus

yes

Open Access

yes

Open Access Type

Gold: This publication is openly available in an open access journal/series

This document is currently not available here.

Share

COinS