Short-term Load Forecasting Based on a Combined ICEEMDAN-PE and IDBO-Informer Model

To address the problems of insufficient noise processing, limited feature extraction ability and complex model training when using traditional methods to deal with complex load data, an innovative forecasting model based on a combined ICEEMDAN-PE and IDBO-Informer is proposed. Firstly, the raw load data were preprocessed using wavelet soft-threshold denoising algorithm to reduce noise interference. Secondly, ICEEMDAN was used for multi-scale decomposition of load data to precisely characterize load features, and the permutation entropy was used to evaluate the component complexity. Finally, an improved Dung Beetle Optimizer (IDBO) was proposed by synergistically integrating chaotic and opposition-based learning strategies for population initialization, incorporating adaptive step size, convex lens opposition imaging, and stochastic differential mutation strategies. This approach optimizes hyperparameters of the Informer forecasting model, significantly enhancing computational efficiency and prediction accuracy. The experimental results show that the model performs well in short-term load forecasting, with MAE of 81.3 MW (the original load data range is about 500 MW to 1500 MW), RMSE of 109.2 MW and R2 score of 0.991, which is much better than the traditional method, and fully verifies the innovation and superiority of the model.