A Study on Feature Dimension Selection for EV Charging Load Forecasting Under Few-Shot Scenarios

Existing research on charging load forecasting generally assumes that integrating multi-dimensional features can improve model accuracy, yet this conclusion presupposes sufficient training data and may not hold in practical scenarios where newly deployed charging piles have scarce data. To address this issue, this paper proposes a two-stage federated transfer learning framework. The FedProx algorithm and LSTM-SelfAttention model are adopted to train global models with single feature and six features respectively, which are then transferred to new stations for few-shot prediction. Experimental results based on actual datasets show that, under 15-day few-shot conditions, the single-feature model reduces RMSE by 27.2% compared with the six-feature model. This conclusion provides direct engineering guidance for feature selection of newly deployed charging piles during the data-scarce phase.