Short-term household power load forecasting based on ICEEMDAN decomposition and CNN-BiLSTM-SA hybrid model

Abstract In recent years, the introduction of the dual carbon strategy has made the development of new power systems more and more crucial. The accurate and efficient prediction of household load has become a hot topic in this field. This paper proposes a hybrid model based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), integrating convolutional neural networks (CNN), bidirectional long short-term memory (BiLSTM), and self-attention (SA) techniques, and incorporating the wavelet domain denoising (WDD) algorithm to enhance data smoothness. The spatial feature extraction capability of CNN and the bidirectional time series extraction capability of BiLSTM are fully utilized, and the ability of SA to highlight the key historical time node information is combined to reduce the loss of historical information. To validate the effectiveness of this method, real household load data is used for verification. Through comprehensive experimental results comparison from multiple perspectives, it is found that both prediction accuracy and error have been significantly improved, demonstrating great potential for this model in addressing increasingly complex new power systems in electrical engineering applications.