MSPatch: A multi-scale patch mixing framework for multivariate time series forecasting

Multivariate time series forecasting is a popular yet challenging topic in fields such as traffic management and weather forecasting, where data often exhibit complex temporal variations. There remains a significant research gap in accurately understanding the correlations between multiple variables in a sequence at different times. Patch pre-processing has shown promise recently, but fixed patch lengths limit its potential. To address this problem, we propose MSPatch (Multi-Scale Patch Mixing for Time Series Forecasting), which decomposes a series into multi-scale patches to capture short-term and long-term patterns. Our main innovations include the Multi-scale Patch Embedding (MSPE) module, Patch Linear Attention (PLA) module, and Hybrid Patch Convolution (HPC) module. This multi-scale patch design allows MSPatch to decompose sequences into different scales and blend seasonal and trend components from fine to coarse scales, thereby flexibly capturing various temporal features. Extensive experiments demonstrate that our model achieves state-of-the-art performance in long-term forecasting tasks, with improvements of 14.5% and 10.0% in prediction accuracy on MSE and MAE, respectively, compared to the benchmark PatchTST. These results highlight MSPatch's ability to flexibly model diverse temporal features, significantly enhancing forecasting accuracy and robustness. By addressing key limitations of existing methods, MSPatch provides a scalable and impactful solution for real-world multivariate time series forecasting challenges.