D-MambaFormer: integrating Mamba and transformer for long-term energy time series forecasting

Time series forecasting in the context of energy necessitates balancing short-term and long-term dependencies for accurate predictions. As the forecast horizon extends, it becomes increasingly challenging for existing methods to model both types of relationships effectively over long sequences. Transformers excel in modeling short-to mid-term dependencies but are often criticized for their quadratic computational cost. On the other hand, Mamba, with its near-linear complexity through a selective state-space mechanism, provides an advantage for long-term forecasting. To combine both effectiveness and efficiency, we propose D-MambaFormer , a hybrid architecture that introduces two key innovations: the Mixture of Patch Extractors , an adaptive method to enhance time series patch representations, and the Mixture of Architectures , which integrates a Mamba-based encoder and a self-attention-based encoder to model multi-level temporal dependencies. Extensive experiments demonstrate that D-MambaFormer achieves strong and consistent forecasting performance across widely used public datasets in clean energy scenarios.