A long-term multivariate time series forecasting network combining series decomposition and convolutional neural networks

Abstract

In multivariate time series forecasting tasks, expanding forecast length and improving forecast efficiency is an urgent need for practical applications. Accurate long-term forecasting of multivariate time series is challenging due to the entangled temporal patterns of multivariate time series and the complex dependencies between variables at different periods. However, it is unreliable for most current models to capture temporal and inter-variable dependencies in intertwined temporal patterns. Furthermore, the Auto-Correlation mechanism cannot precisely capture the local dynamics and long-term dependencies of time series. To address these issues, we propose a concise and efficient model named SDCNet, which integrates time series decomposition and convolutional neural networks (CNNs) into a unified framework. Unlike previous approaches, SDCNet untangles the entangled temporal patterns and uses CNNs to capture the dependencies in both temporal and feature dimensions, respectively. Specifically, SDCNet progressively decomposes seasonal and trend-cyclical components from past time series, and uses temporal and feature convolution modules to extract seasonal patterns and inter-variable dependencies, respectively. Compared to competing methods, SDCNet achieves the best performance on all of four real-world datasets with a relative accuracy improvement of 16.73%. In addition, SDCNet achieves a relative performance gain of 23.87% on datasets with no significant periodicity.