DTMamba: Dual Twin Mamba for Time Series Forecasting

Long-term Time Series Forecasting (LTSF) has always been an important task where models need to effectively capture hidden patterns in the time series in order to make accurate predictions about future states. The State-Of-The-Art (SOTA) models in this area are mostly based on Transformers, but the prediction accuracy still needs to be improved. Recently, Mamba has emerged as a promising approach for modeling sequential data, especially for autoregressive tasks with long sequences. Therefore, in this paper, we propose an LTSF model called DTMamba. DTMamba utilizes innovative dual twin Mamba blocks to extract long-term dependencies in time series, while also incorporating residual network to enhance the overall predictive capability. We perform experiments with 8 publicly available datasets and compare DTMamba with 11 SOTA models. The experimental results show that DTMamba outperforms the SOTA models in terms of performance. In particular, our proposal has obvious advantages when dealing with low-dimensional data and predicting longer time series.