🤖 AI Summary
This work addresses the challenge of modeling highly coupled trend, multi-scale seasonal, and irregular residual components in multivariate time series forecasting, where cross-variable shared structures are difficult to capture effectively. To this end, the authors propose DecompSSM, a novel framework that employs three parallel deep state space model branches to explicitly learn trend, seasonality, and residual components. The approach integrates a learnable decomposition mechanism, an input-dependent adaptive temporal scale predictor, a cross-variable context refinement module, and an orthogonality-constrained auxiliary loss to enable end-to-end structure-aware disentangled modeling. Extensive experiments demonstrate that DecompSSM significantly outperforms strong baselines on benchmark datasets including ECL, Weather, ETTm2, and PEMS04, validating the effectiveness of component-wise state space modeling combined with global context refinement.
📝 Abstract
Multivariate time series (MTS) forecasting is crucial for decision-making in domains such as weather, energy, and finance. It remains challenging because real-world sequences intertwine slow trends, multi-rate seasonalities, and irregular residuals. Existing methods often rely on rigid, hand-crafted decompositions or generic end-to-end architectures that entangle components and underuse structure shared across variables. To address these limitations, we propose DecompSSM, an end-to-end decomposition framework using three parallel deep state space model branches to capture trend, seasonal, and residual components. The model features adaptive temporal scales via an input-dependent predictor, a refinement module for shared cross-variable context, and an auxiliary loss that enforces reconstruction and orthogonality. Across standard benchmarks (ECL, Weather, ETTm2, and PEMS04), DecompSSM outperformed strong baselines, indicating the effectiveness of combining component-wise deep state space models and global context refinement.