New perspectives on multivariate time series forecasting: Lightweight networks combined with multi-scale hybrid state space models

Abstract

In the real world, applications such as industrial energy planning and urban transport planning require forecasting future trends from historical data. Due to the significance and complexity of these issues, there is an urgent need for robust prediction algorithms that can handle long-term time series forecasting. In recent years, transformer-based algorithms have emerged and demonstrated great potential. However, their computational costs are substantial, leading to inefficiency. A lightweight module called LSM is proposed to enhance the accuracy of Long-term Time Series Forecasting (LTSF). This model exhibits linear scalability and low computational costs. By effectively combining deep learning models with a hybrid state space model architecture, it efficiently captures dependencies at different scales within patches to predict global and local contexts accurately. Additionally, to further improve algorithm performance and computational efficiency, this model adopts a “strong encoder-light decoder” architecture design. Experimental results on 8 benchmark datasets demonstrate that LSM performs exceptionally well in long sequence prediction tasks by exhibiting strong robustness and effectiveness compared to State-Of-The-Art approaches (SOTA). Moreover, LSM significantly enhances accuracy while reducing computational requirements. Code availability: https://github.com/hao-fei-hub/LSM/.