DC-Mamber: A Dual Channel Prediction Model Based on Mamba and Linear Transformer for Multivariate Time Series Forecasting

Multivariate time series forecasting (MTSF) requires capturing both local intra-variable patterns and global cross-timestep dependencies. Existing data processing strategies are divided into channel-independent, focusing on each variable’s local temporal features, and channel-mixing, emphasizing global temporal dependencies. Mainstream models are based on Transformer and the emerging Mamba. Transformers effectively model global dependencies but face quadratic complexity and weak local sensitivity, while Mamba, based on state space models (SSMs), achieves linear complexity and efficient long-range modeling yet struggles with parallel global contextual information. To address these limitations, we propose DC-Mamber, a dual-channel prediction model based on Mamba and linear Transformer. Specifically, DC-Mamber employs a Dual-Channel Embedding Layer to map inputs into two streams reflecting distinct strategies, which are processed separately by a Dual-Channel Encoder and subsequently integrates the outputs via a Feature-fusion module. Experiments on four datasets demonstrate DC-Mamber’s superior accuracy over existing models.