Electric Vehicle Charging Demand Prediction in Shenzhen Based on Random Forest and LSTM

Abstract

Reliably forecasting how much energy electric vehicles (EVs) will require at public chargers is a prerequisite for infrastructure planning and matching available capacity to real-world usage. This study compares two representative approaches, Random Forest and Long Short-Term Memory (LSTM), to forecast hourly charging load, drawing on the UrbanEV dataset that spans 275 charging zones across Shenzhen, China. The Random Forest model incorporates 1 -hour and 24-hour lag features and ensemble learning, while the LSTM model leverages a sliding window to capture temporal dependencies. Evaluation on the test set (February 2023) shows that Random Forest achieves an RMSE of 606.74 kWh, MAE of 465.07 kWh, and MAPE of 8.32%, outperforming LSTM by approximately 24-25% across all metrics. The advantage of Random Forest is attributed to the explicit lag features encoding daily periodicity, higher data efficiency with limited training samples, and the regularity of the single-station charging pattern. These findings suggest that well-designed feature engineering combined with ensemble methods can outperform deep learning models on moderate-sized datasets with strong periodic patterns.