Forecasting residential EV charging load using enhanced decomposition and Conv-Informer networks

Abstract Accurately predicting the charging load of residential electric vehicles (EVs) is crucial for optimizing the power grid and infrastructure planning. In order to reduce the impact of high fluctuation of charging load on prediction results, this paper proposes a hybrid prediction model that integrates enhanced decomposition technology and Conv-Informer network. Firstly, identify key features through Pearson correlation analysis. Then, the fully integrated empirical mode decomposition (CEEMDAN) with adaptive noise is used to perform initial decomposition on the charging load data, obtaining multiple intrinsic mode functions (IMFs). These IMFs are divided into high, medium, and low frequency components based on sample entropy, with the high frequency components further decomposed using variational mode decomposition (VMD) and their modulus optimized using the center frequency method. Input all components into the Conv-Informer model for prediction, where the convolutional layer extracts local load features, the Informer module captures long-term load dependencies, and finally weights and superimposes the predicted values of each subsequence to obtain the final result. Based on real data validation experiments in a residential area in Zhejiang Province, China, the proposed model achieved an RMSE of 5.74 kW and a MAE of 4.03 kW. Compared with other common models, its prediction error has been reduced by about 30% (e.g. LSTM: RMSE = 9.28 kW, MAE = 6.64 kW; Single Informer: RMSE = 7.61 kW, MAE = 4.54 kW).It can be seen that the method proposed in this article has good accuracy and stability, and is suitable for predicting the charging load of residential electric vehicles.