Preactivated Residual Neural Network With Long Short-Term Memory to Predict EV Charging Demand at an Individual Fast-Charging Station

Abstract

This study proposes a preactivated residual neural network (ResNet) with long short-term memory (LSTM) to predict electric vehicle (EV) charging demand at an individual fast-charging station. While fast-charging stations offer convenience to EV users, the use of fast-charging stations can also threaten the stability and quality of the power system. Therefore, it is important to accurately forecast the charging demand at individual fast-charging stations for the operation of the power system. The proposed model incorporates two deep learning models: ResNet and LSTM. The ResNet is used to perform the feature extraction needed for forecasting fast-charging patterns. The LSTM performs forecasting of fast-charging demand based on sequential input. The proposed model ensures superior forecasting performance without vanishing gradient. Furthermore, the structure of the preactivated ResNet enables optimal parameter updates based on the loss function of mean squared error (MSE). The proposed model was evaluated with real-world data from EV fast-charging stations in Jeju Island, South Korea. The maximum prediction performance of the proposed model was attained with 8.04% in the normalized root MSE and a mean absolute error (MAE) of 4.71 kW.