Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DTMamba : Dual Twin Mamba for Time Series Forecasting

Zexue Wu, Yifeng Gong, Aoqian Zhang

TL;DR

DTMamba addresses the challenge of long-term multivariate time series forecasting by extending the Mamba framework with RevIN normalization, Channel Independence, and dual TMamba blocks. The method integrates Embedding layers, residual connections, dropout, and a projection layer to produce accurate forecasts, while ensuring scalable computation through a novel twin Mamba architecture. Empirical results across six real-world datasets and multiple baselines demonstrate state-of-the-art performance and robustness, with ablations confirming the value of channel independence and residuals. The work offers a practical, scalable approach for high-quality multivariate LTSF with potential impact on finance, energy, weather, and more.

Abstract

We utilized the Mamba model for time series data prediction tasks, and the experimental results indicate that our model performs well.

DTMamba : Dual Twin Mamba for Time Series Forecasting

TL;DR

DTMamba addresses the challenge of long-term multivariate time series forecasting by extending the Mamba framework with RevIN normalization, Channel Independence, and dual TMamba blocks. The method integrates Embedding layers, residual connections, dropout, and a projection layer to produce accurate forecasts, while ensuring scalable computation through a novel twin Mamba architecture. Empirical results across six real-world datasets and multiple baselines demonstrate state-of-the-art performance and robustness, with ablations confirming the value of channel independence and residuals. The work offers a practical, scalable approach for high-quality multivariate LTSF with potential impact on finance, energy, weather, and more.

Abstract

We utilized the Mamba model for time series data prediction tasks, and the experimental results indicate that our model performs well.
Paper Structure (37 sections, 4 equations, 8 figures, 5 tables, 1 algorithm)

This paper contains 37 sections, 4 equations, 8 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional methods
  4. RNN-based methods
  5. Transformer-based methods
  6. Linear-based methods
  7. TCN-based methods
  8. Applications of Mamba
  9. Proposed Method
  10. Problem Statement
  11. Normalization
  12. Channel Independence & Reversed Channel Independence
  13. Twin Mamba
  14. Embedding Layers
  15. Residual
  16. ...and 22 more sections

Figures (8)

  • Figure 1: Overall structure of DTMamba.
  • Figure 2: Details of TMamba Block.
  • Figure 4: Sensitivity of Dropouts
  • Figure 5: Sensitivity of Linear's Parameters
  • Figure 6: Sensitive of Mamba Dimension Expansion Factor
  • ...and 3 more figures

Theorems & Definitions (1)

  • definition 1: Multivariate Time Series