DBformer: A Dual-Branch Temporal Prediction Network Based on Transformer

Multivariate time series forecasting plays a crucial role in various domains, including energy, transportation, and climate science. However, such data often exhibit complex inter-dependencies that traditional modeling approaches struggle to capture effectively. While Transformer-based time series forecasting models have shown promising performance, they primarily focus on temporal dependencies and rely on point-wise self-attention mechanisms to capture inter-variable relationships. This approach is insufficient for modeling the intricate cross-correlations among variables. To address this issue, we propose DBformer, a dual-branch Transformer-based network for time series forecasting. After preprocessing, the input sequence is fed into two parallel branches: a global branch, where multiple variables are stacked together and processed through global attention to capture inter-variable relationships, and a local branch, where channel-independent and patch-based mechanisms are employed to focus on local sequence details. The extracted sequence features from both branches are then adaptively fused through a spatiotemporal feature fusion module. This design effectively captures and models both inter-variable dependencies and temporal correlations, thereby enhancing forecasting accuracy. Extensive experiments on multiple real-world datasets demonstrate that DBformer outperforms advanced Transformer models, achieving a 1.3% improvement in Mean Squared Error (MSE) and a 3.4% improvement in Mean Absolute Error (MAE). Additionally, compared to the best-performing Convolutional Neural Network-based (CNN-based) model, our method achieves a relative reduction of 11.5% in MSE and 7.7% in MAE.