MSRNet: Multi-level series decomposition and stepwise reconstruction network for load forecasting

Linear models and patch-based Transformer models have made new advances in the field of time series forecasting. However, power load sequences present nonlinear structures and complex trends, making prediction challenging. Additionally, dynamic correlations between load sequences and external features, like environmental and social factors, are difficult to model effectively. This paper introduces MSRNet, a Multi-level Series Decomposition and Stepwise Reconstruction Network, which enhances load forecasting by leveraging external features. We found that existing methods for multivariate correlation modeling are not only ineffective but also lack interpretability. Our approach gradually separates the trends of different scales in the load sequence and external features through multi-level seasonal-trend decomposition. Subsequently, a graph-enhanced learning module constructs multivariate correlations for these trends at different scales. Meanwhile, the multi-scale Transformer module captures frequency variations in the seasonal components of the load sequences. By stepwise reconstruction of the trend and seasonal components, the final forecasting result is obtained. Extensive experiments were conducted on six real-world datasets. Results indicate our model achieves more accurate predictions compared to the latest baselines. Particularly significant improvements were observed on datasets containing external features, reducing MSE by 53.66% and MAE by 31.62%. Code is available at this repository: https://github.com/Del-Vecchio/MSRNet .