Reconstructing Temporal Heterogeneity: A Multidomain Collaborative Analysis Framework for Robust Time-Series Forecasting

Real-world time-series data typically exhibit multiscale patterns, heterogeneous structures, and nonstationarity. Capturing these multifaceted temporal variations remains a fundamental yet challenging problem in time-series forecasting. Despite considerable advancements in recent years, most existing approaches still depend on conventional decomposition paradigms that separate time series into trend and seasonal components, while neglecting abrupt events and systemic perturbations in practical applications. This limitation not only introduces forecasting bias but also hinders model generalization and interpretability in complex environments. To tackle these difficulties, we propose a multidomain collaborative analysis framework (MCAF). The framework addresses temporal heterogeneity through structural decoupling and domain-specific modeling. Specifically, each time series is decomposed into three distinct components: trend, seasonal, and residual. These components are then processed through specialized modeling pathways, each tailored to capture their unique dynamics in either the time, frequency, or time-frequency domain. A unified fusion mechanism subsequently integrates the extracted multidomain representations. Extensive experiments show that MCAF consistently delivers high accuracy and robustness across diverse forecasting horizons. The source code is available at https://github.com/hengruili1997/MCAF_TSF/tree/main