PatchRNN: A Multi-Scale Decoupled and Structure-Aware Framework for Long-Term Time Series Forecasting

Long-term time series forecasting is important in energy, transportation, and meteorology, yet existing methods struggle to model multi-scale features and preserve local sequence structure. We propose PatchRNN, a decomposition- and structure-aware framework for multivariate forecasting. Its parallel trend branch uses multi-scale patch embedding to capture and fuse long-term patterns at different temporal granularities. The seasonal forecasting branch separates the residual components via average pooling and employs a segmented recurrent neural network (SegRNN) to efficiently model their periodic fluctuations. Joint optimization combines point-wise error with Patch-wise Structural Loss (PSLoss). Fourier-based adaptive partitioning enables PSLoss to constrain local trend correlation, fluctuation variance, and mean. Extensive experiments across multiple real-world benchmark datasets demonstrate that our model obtains competitive long-term forecasting performance while preserving favorable local-structure consistency. On the real-world ocean-wave dataset School.nc, our model yields an MSE of 0.814 at prediction horizon 96, corresponding to an 8.64% reduction over the best competing baseline.