MDEformer: A spatio-temporal decoupling transformer with the multidimensional information encoding for certain traffic flow prediction

With the rapid advancement of Intelligent Transportation Systems (ITS), accurate traffic flow prediction has become a critical challenge. Although existing deep learning models are capable of capturing the spatio-temporal dependencies in traffic data to some extent, they still face limitations in modeling spatio-temporal features, long-term and short-term temporal dependencies, and the dynamic-static spatial relationships at local and global scales. This paper proposes a Spatio-Temporal Decoupling Transformer with Multidimensional Information Encoding (MDEformer) to address these issues. MDEformer effectively captures spatio-temporal features through the integration of multidimensional information encoding. Moreover, the model adopts a decoupled design of temporal and spatial encoder layers. In the temporal encoder layer, a GRU is employed to replace the linear mapping in the multi-head self-attention mechanism, thereby facilitating improved capture of short-term fluctuations and long-term trends. In the spatial encoder layer, we integrate the graph fusion module with Chebyshev graph convolution to replace the conventional mapping in the multi-head self-attention mechanism, thereby enhancing the capability to model local and global dynamic-static spatial dependencies. Experiments on four real-world traffic datasets demonstrate that MDEformer significantly outperforms other baseline methods in terms of prediction accuracy.