Answers are grounded in this paper and cite their sources.
Accurate prediction of EV charging demand is fundamental to grid stability, real-time pricing, and charging station management, but few prediction methods are able to utilize more than just historical demand, and are able to reveal why a particular forecast was made. In this paper we develop and experimentally investigate an explainable multi-modal forecasting approach to hourly EV charging demand, which fuses charging session data, weather information, grid prices, and grid operating information (voltage, frequency, transformer load) with station occupancy and traffic signal status. Using 15,304 charging session records from 55 charging stations between January 2019 and July 2020, we construct a 13,680-hour multi-modal data panel including calendar features, lags, and rolling statistics. We evaluate five forecasting architectures, including linear regression, Random Forest, XGBoost, a Long Short-Term Memory network, and our proposed attention-based Transformer network with a learned temporal pooling mechanism to interpret attention. XGBoost and Random Forest are the two strongest models overall and slightly outperform each other on different metrics: XGBoost achieves the lowest RMSE and highest R² (RMSE = 5.63 kWh, R² = 0.980), while Random Forest achieves a marginally lower MAE and MAPE (MAE = 1.67 kWh, MAPE = 7.39%, versus XGBoost’s MAE = 1.72 kWh, MAPE = 8.36%); a paired significance test shows this MAE difference is not statistically significant, so the two tree ensembles are best described as practically comparable, with XGBoost’s edge concentrated in the largest-error hours. Both substantially outperform our attention-based Transformer network (RMSE = 27.98 kWh, R² = 0.513) and LSTM network (RMSE = 30.02 kWh, R² = 0.440) at current data scale; in contrast, the attention weights from the transformer provide an intuitive explanation on which past hours are the most influential ones in forecasting the current charging demand. SHAP value analysis of the best model identifies the recent past charging demand, transformer loading, and the hour of the day as the most significant predictors, and ablation tests confirm a 4–6% increase in RMSE from the removal of grid and calendar data, whereas a negligible impact is observed from the removal of station occupancy/traffic features. All differences in forecasting performance across various architectures are confirmed by Diebold-Mariano tests to be statistically significant, and a sensitivity analysis to model hyperparameters reveals that the Transformer’s performance is largely independent of the exact model parameter count over a 6.5x range, indicating data volume rather than model complexity is currently the bottleneck. Further analysis of forecasting errors demonstrates that both the best tree-based model and the Transformer fail at the same sudden morning weekday charging demand surges. An alternative variant of the Transformer network using modality-attention fusion (learning weights for each input feature category in a data-dependent way rather than simple concatenation) slightly but significantly reduces the mean absolute error (but not the RMSE) compared to the non-modality-attention approach and offers learned modality weights that confirm the relative importance ordering from the ablation study. We conclude that for aggregated single-station charging demand forecasting, tree-based models with SHAP are both the most accurate and explainable, whereas attention-based models provide a promising approach for integrated feature explanation with greater scalability potential as multi-year and multi-station data become available. We discuss the implications for smart-grid operation and a road map for widespread spatial deployment of our model.
No takes yet. Share an insight, caveat, or question.
Mushfiq Us Salehin (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: