Grid load forecasting based on hybrid ensemble empirical mode decomposition and CNN–BiLSTM neural network approach

Abstract This article proposes an amalgamation of ensemble empirical mode decomposition (EEMD) and the convolutional neural network–bidirectional long short-term memory (CNN–BiLSTM) for the prediction of electricity grid load. Initially, the original load time series undergoes decomposition using EEMD, resulting in intrinsic mode functions (IMFs) that capture various load characteristics. Subsequently, a correlation analysis selects several IMFs closely related to the original sequence. These chosen IMFs are then utilized as input, with separate application of a one-dimensional CNN and a BiLSTM model for modeling and prediction purposes. The CNN automatically extracts temporal features from the different IMFs via its convolutional layers, whereas the BiLSTM effectively captures both short-term and long-term dependencies. In the end, a linear combination is employed to integrate the IMF predictions and reconstruct the final forecast for the electricity grid load. Experimental results demonstrate that this hybrid integration model, combining the adaptive decomposition ability of EEMD, feature extraction capability of CNN and temporal modeling ability of BiLSTM, improves the accuracy and robustness of electricity grid load forecasting compared to single models and ensemble models without EEMD.