Traffic flow prediction based on MSCNN and attention mechanism

Aiming at the drawbacks of current traffic network flow prediction methods, such as low capability to explore the potential trend of spatial data and ignoring the long-term data correlation, this paper proposes a novel traffic flow prediction model based on MSCNN and attention mechanism. The model designs an RMSCNN module for considering the spatio-temporal dependence of long and short-term traffic flows, in which a residual block is introduced to alleviate the problem of excessive downsampling loss during feature extraction. In addition, a temporal multi-attention module based on the Transformer framework is used to highlight long time series features. To validate the method, traffic flow prediction experiments were conducted using the PEMS04 dataset. The results show that the RMSCNN-Trans prediction model reduces the Root Mean Square Error (RMSE) by at least 0.14, the Mean Absolute Error (MAE) by at least 1.33 and the Mean Absolute Percentage Error (MAPE) by at least 0.02 compared to several commonly used models.