Multi-Task Learning for Short-Term Electric Vehicle Charging Demand Forecasting: Joint Prediction of EV Count and Energy Consumption

The rapid adoption of electric vehicles (EVs) is transforming the transportation sector, but it also presents significant challenges to power systems, particularly in the context of dynamic and fluctuating electricity demand for EV charging. Accurate short-term forecasting of EV charging demand, including both the number of EVs at charging stations (EV count) and the energy consumption required for charging, is critical to managing grid stability and efficiency. Traditional forecasting models treat these two tasks independently, failing to capture the inherent interdependencies between them. To address this, we propose a Multi-Task Learning (MTL) framework that jointly predicts EV count and energy consumption. By leveraging shared representations across these tasks, the MTL model improves forecasting accuracy and overcomes the limitations of single-task models. We further enhance the model by incorporating additional spatiotemporal features to capture the cyclical and geographical nature of EV charging demand. Extensive experiments across six real-world datasets and multiple model backbones demonstrate that our approach consistently outperforms baseline methods. For instance, on the JPL dataset, applying MTL to the LSTM model reduces the energy prediction RMSE by 46.96%, MAE by 64.18%, and improves CORR by 69.21% compared to its single-task counterpart.