Explainable AI Model Forecasts EV Charging Demand
Tree-based models outperformed deep learning approaches in predicting hourly charging demand across 55 stations.
Updated on Nov. 13, 2026 in Electric Vehicles

Researchers developed an explainable multi-modal forecasting approach to hourly EV charging demand, testing five architectures on 15,304 charging session records from 55 stations. Tree-based models outperformed the attention-based Transformer and LSTM networks on forecasting accuracy.
Why it matters
Accurate charging demand prediction supports grid stability, real-time pricing, and charging station management as electric vehicle use grows. The study's use of SHAP analysis also makes the model's predictions more interpretable to operators.
XGBoost recorded an RMSE of 5.63 kWh and an R² of 0.980 across a 13,680-hour multi-modal panel. Removing grid and calendar data increased RMSE by 4-6%, though performance on data beyond the July 2020 study window is not reported.
The details
The model fuses charging session data with weather, grid prices, grid operating information, station occupancy, and traffic signal status. Model explanation is provided through SHAP value analysis and transformer attention weights, which identified recent past charging demand, transformer loading, and hour of day as the most significant predictors.
Timeline
January 2019 to July 2020: Period of charging session data collection.
November 13, 2026: Article published as accepted early-access research.
Roadmap
The result adds to a growing body of evidence that gradient-boosted tree models can match or outperform LSTM and Transformer networks on structured tabular time-series data like charging session records. Its emphasis on SHAP-based explanation also reflects a broader push to make forecasting models usable by grid and charging-network operators who need to justify their decisions.
Better hourly demand forecasts can help charging networks manage peak loads and set pricing that reduces congestion for EV drivers. Operators that adopt interpretable models can also explain why certain hours are expected to be busy.
The takeaway
Charging operators should weigh interpretable tree-based forecasting models alongside deep learning options, since the study found the simpler approach more accurate on this dataset. Testing any forecasting tool against recent local session data before deployment remains the most practical step.
Further reading
For broader context on how charging demand is changing, explore more coverage in Electric Vehicles.



