10 July 2025 in IEEE Transactions on Industrial Electronics
Existing time series prediction models struggle with long-sequence data due to limitations in capturing long-range dependencies and processing multichannel information simultaneously. To address this gap, we first propose the channel independence bidirectional gated Mamba (CIBG-Mamba), a novel model enhancing gated recurrent neural networks by integrating Mamba into channel-independent recurrent units. At its core, CIBG-Mamba processes each data channel independently, with each channel utilizing three Mamba-enhanced recurrent units. This design improves memory capacity while enabling dynamic interchannel exchange, allowing the model to effectively handle multidimensional time series (MTS) data. Furthermore, an interactive recurrent mechanism (IRM) is introduced to optimize bidirectional information handling, employing channel-specific adaptive state transition functions. IRM significantly enhances the CIBG-Mamba's ability to capture complex time series dynamics across multiple dimensions. Finally, by fusing the above technologies, the prediction approach, i.e., the channel independence bidirectional gated Mamba with interactive recurrent mechanism (CIBG-Mamba-IRM) is constructed for time series forecasting. Additionally, a dataset with one year of data from five new energy vehicle air conditioners (NEVAC) is created. Through experiments on this dataset and five public time series datasets, our method proved effective in learning long-term dependencies and handling MTS data, achieving state-of-the-art results. Our code is available at https://github.com/XYouZheng/CIBG-Mamba.
24 June 2026 in Expert Systems with Applications