Prediction of EV Charging Load Using Two-Stage Time Series Decomposition and DeepBiLSTM Model

Abstract

The fast adoption of electric vehicles (EVs) has resulted in a growing concern for the planning and operation of the distribution power system. Thus, it is crucial to comprehensively assess the influence of EV charging load on the distribution system. To mitigate the operational risks and optimize control measures of the power network, a two-stage time series decomposition and BiLSTM deep learning architecture is proposed for accurately predicting EV charging load series. In the first stage, a Variational Mode Decomposition (VMD) evaluation method is introduced for assessing the decomposition loss, and the Lion Swarm Optimization (LSO) is employed to determine the optimal combination of decomposition parameters, thereby reducing the empirical parameter randomness and minimizing the signal loss during the decomposition process. Then, Bidirectional Short and Long-term Memory Network model (BiLSTM) is established for forecasting each sub-sequence obtained through the decomposition process. In the second stage, a DeepBiLSTM model is built to realize the accurate prediction of residual series from the first stage. The performance of the proposed architecture was tested using real-world charging load data collected from EV stations located in Southern China. It demonstrates that our proposed forecasting method can outperform traditional algorithms in terms of MAPE and RMSE metrics for 1h-ahead, 2h-ahead, 4h-ahead, and 24h-ahead EV load prediction.