Efficient Multivariate Time Series Forecasting with SC-TWRNet: Combining Adaptive Multi-Resolution Wavelet and Parallelizable Decomposition
Year: 2026
Type: article
Abstract: Long-term multivariate time series forecasting serves as a fundamental analytical tool across diverse domains, such as energy management, transportation analysis, and meteorology. However, conventional modeling paradigms often yield suboptimal results as they fail to adequately capture non-stationarity and multi-scale temporal correlations. While frequency-domain methods offer theoretical clarity, representative efficient spectral-domain architectures often rely on magnitude-based spectral pruning to ensure efficiency, inadvertently discarding high-frequency transient signals essential for non-stationary forecasting. To address these limitations, we propose the Structural Component-based Temporal Wavelet-Refine Network (SC-TWRNet), a framework that orchestrates adaptive wavelet filtering with explicit structural temporal decomposition. The architecture is anchored by the Adaptive Multi-Resolution Wavelet (AMRW) filter, designed to generate time-frequency representations while maintaining linear computational complexity. Concurrently, a structural temporal decomposition module decouples the input stream into distinct trend, seasonal, and residual components for targeted modeling. Extensive experiments on eight standard datasets demonstrate that SC-TWRNet achieves superior predictive accuracy compared to state-of-the-art baselines while maintaining linear computational complexity for efficient high-dimensional modeling.
Source: Algorithms
DOI: 10.3390/a19020155
Authors: Yu Chen, Hanshen Li
Institution: Dalian University
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,930
OpenAlex ID: W7128953431
FWCI:
Cites:
Cited by:
Related to: