Fractional-order Difference Decoupling and Pattern Extraction for Multivariate Time Series Forecasting

Multivariate time series forecasting is essential in many real-world applications. However, inherent challenges such as non-stationarity, multi-scale temporal patterns, weak inter-variable correlations, and long-range dependencies limit the effectiveness of existing models. The modeling of heterogeneous temporal dynamics in a unified space, which often leads to structural interference and degraded leaning efficiency. This study proposes a novel forecasting framework, termed the Fractional Differencing Decoupling and Pattern Extraction Model (FDDPEM). Built upon a structural decoupling–reconstruction paradigm, FDDPEM first separates trend and seasonal components using fractional differencing, which stabilizes the series while preserving longmemory characteristics. Each component is then modeled with specialized mechanisms through multi-scale segmentation, clustering, and two-dimensional convolution. Extensive experiments on multiple benchmark datasets demonstrate that FDDPEM consistently outperforms state-ofthe-art methods, particularly in handling weakly seasonal data and complex inter-variable relationships. Ablation studies further validate the rationality and effectiveness of the proposed structural design.