A Decomposition Modeling Framework for Seasonal Time-Series Forecasting

TL;DR

To address seasonal time-series forecasting with complex long-range dependencies, the paper introduces MSSD, a decomposition framework that splits a univariate series into $x_u$, $x_p$, and $x_d$ within each cycle and predicts each component separately before aggregating to the final forecast $y_t$. Ascending and Descending components are modeled via simple linear regression for efficiency and interpretability, while Peak is captured by SDNet—a multi-scale convolutional network using Conv2d and dilated causal convolutions to learn local/global peak fluctuations. Across CAISO, Electricity, and Traffic datasets, MSSD achieves around a 10% reduction in error compared to strong baselines for both short- and long-horizon forecasts in univariate and multivariate settings, demonstrating superior forecasting accuracy. The framework also emphasizes interpretability through explicit phase decomposition and shows robustness and efficiency advantages, indicating practical impact for real-world seasonal forecasting. Overall, MSSD provides a scalable, interpretable, and effective approach for seasonal time-series prediction that outperforms transformer- and CNN-based baselines on multiple real-world datasets.

Abstract

Seasonal time series exhibit intricate long-term dependencies, posing a significant challenge for accurate future prediction. This paper introduces the Multi-scale Seasonal Decomposition Model (MSSD) for seasonal time-series forecasting. Initially, leveraging the inherent periodicity of seasonal time series, we decompose the univariate time series into three primary components: Ascending, Peak, and Descending. This decomposition approach enhances the capture of periodic features. By addressing the limitations of existing time-series modeling methods, particularly in modeling the Peak component, this research proposes a multi-scale network structure designed to effectively capture various potential peak fluctuation patterns in the Peak component. This study integrates Conv2d and Temporal Convolutional Networks to concurrently capture global and local features. Furthermore, we incorporate multi-scale reshaping to augment the modeling capacity for peak fluctuation patterns. The proposed methodology undergoes validation using three publicly accessible seasonal datasets. Notably, in both short-term and long-term fore-casting tasks, our approach exhibits a 10$\%$ reduction in error compared to the baseline models.

Figures (10)

• Figure 1: Overall architecture of MSSD.
• Figure 2: the specifics of the CAISO dataset. (1) represents the CAISO dataset itself; (2) delineates the windows for input and output series; and (3) showcases the decomposition results of CAISO.
• Figure 3: Seasonal component prediction module.
• Figure 4: Local-Global convolution block architecture.
• Figure 5: Efficiency Analysis(local-global module vs self-attention, auto-correlation).