A hybrid electric vehicle load classification and forecasting approach based on GBDT algorithm and temporal convolutional network

Abstract

Due to the participation of large-scale electric vehicles (EVs) in Vehicle-to-Grid (V2G) services, V2G dispatch centers need to predict the charging and discharging (C&D) loads of electric vehicles in a short time period. This study proposes a novel machine learning based approach for EV load forecasting in power supply systems facing high resource uncertainty. This method takes advantage of both Gradient Boosting Decision Tree (GBDT) algorithm and Time Convolutional Network (TCN) model. This study considers the service decisions of EV users and uses the GBDT algorithm to classify the EV discharge load dataset, with 92% accuracy. Also, the TCN model is used to capture local temporal features and predict the future C&D loads. In comparison with other baseline models, e.g. CNN-BILSTM, LSTM, PSO-BP, the stability of the TCN model is superior in real data charging load forecasting, and it is shown that the TCN model has the smallest error. The feasibility of the proposed GBDT-TCN hybrid model is verified by numerical cases,and achieves the RMSE of discharging forecasting less than 6.23%.