Recently, Transformer-based models have achieved remarkable progress in time series forecasting, yet they still suffer from limitations in efficiency and scalability due to quadratic complexity. In contrast, the newly proposed Mamba model demonstrates strong potential with its linear-time complexity and selective state space mechanism, making it more suitable for long-sequence modeling. Meanwhile, convolutional neural networks (CNNs) have long been recognized for their effectiveness in capturing local dependencies and finegrained temporal patterns. To integrate the complementary advantages of these two paradigms, we propose PaDuM, a Patch-Based Dual-Stream Network that jointly leverages CNN and Mamba. Specifically, PaDuM first applies an exponential moving average (EMA) to adaptively decouple input sequences into trend and seasonal components, ensuring better interpretability and reduced noise. Each component is then patchified and modeled through dual streams, where CNN focuses on local seasonal variations while Mamba captures global trends with efficiency. To further enhance stability, we introduce a novel Sigmoid-based weight decay loss, which emphasizes recent predictions while preventing overfitting to distant horizons. Extensive experiments conducted on eight diverse real-world datasets spanning electricity consumption, traffic flow, and meteorological data consistently demonstrate that PaDuM achieves state-of-the-art forecasting performance, while maintaining strong robustness, scalability, and generalization ability across domains. The implementation and resources are publicly available at https://github.com/T-DXVN/PaDuM.
Tao et al. (Sun,) studied this question.