In response to the problem of congestion delay dependence in existing traffic flow prediction models, a dynamic correction method for multi-dimensional long-term time series prediction of traffic flow (MSD-Transformer) integrating multi-scale dual Transformers is proposed. Firstly, a multi-scale feature extraction network is constructed using weight sharing Swin Tiny, which hierarchically mines multi-dimensional traffic temporal features such as short-term fluctuations, daily cycles, and weekly cycles. A bidirectional feature pyramid BiFPN is introduced to build a bidirectional feature fusion pathway, and an adaptive weighting mechanism is used to achieve complementary interaction between high and low resolution features; Secondly, design a spatial residual attention module that combines layered residual attention scores to synchronize sequence self-attention and cross sequence cross attention modeling, accurately characterizing the differential congestion propagation delay of main and branch roads; Thirdly, based on the dual Transformer architecture, a multi-scale spatiotemporal encoding and decoding unit is built, and the model training optimization is completed by combining weighted cross entropy and focus loss fusion function. The experimental results show that compared with mainstream baseline models such as STGCN, DCRNN, and STTN, the proposed model significantly reduces MAE and RMSE indicators in the 15-60 minute full time prediction task; Experimental validation of multi-scale feature extraction network BiFPN, The three major modules of spatial residual attention all have irreplaceable gains. Research has shown that the proposed method can finely model traffic spatiotemporal delays and non-stationary sudden changes in flow, effectively alleviating the performance degradation of long-term time series prediction and providing precise support for dynamic flow regulation of urban road networks.
