Short-term load forecasting based on multi-frequency sequence feature analysis and multi-point modified FEDformer

Given the complexity and dynamic nature of short-term load sequence data, coupled with prevalent errors in traditional forecasting methods, this study introduces a novel approach for short-term load forecasting. The method integrates multi-frequency sequence feature analysis and multi-point correction using the FEDformer model. Initially, variational mode decomposition (VMD) technology decomposes the load sequence into multiple subsequences, each exhibiting distinct frequency characteristics. Subsequently, for each frequency band of the load sequence, the LightGBM algorithm quantifies the correlation between the load and various influencing factors. The filtered features are then input into the FEDformer model, providing preliminary short-term and long-term sequence prediction results. Finally, a point-by-point forecasting method based on a tree model generates multi-point load prediction results by training multiple LightGBM models. Throughout the forecasting process, a weighted threshold α is set, and a hybrid weighting method is utilized to combine the forecast results from different models, culminating in the final short-term load forecast results. Validation of the proposed hybrid model was conducted on an actual dataset from a specific area, The results exhibit higher prediction accuracy, affirming the proposed method as a novel and effective approach for short-term load forecasting.