Urban traffic flow forecasting is of great significance for traffic management and public safety. However, the correlations of traffic raster flow change with time. There are global spatio-temporal correlations in the city, and the contributions of channel-wise features vary on each city region. To tackle these challenges and make more accurate prediction, a novel spatio-temporal neural network model, named 3D-CANet (three-dimensional channel-wise attention network), was designed. A 3D-InnerCA (three-dimensional inner-channel attention) unit was proposed to dynamically capture the global spatio-temporal correlations for different channel-wise features. Meanwhile, an InterCA (inter-channel attention) unit was designed to adaptively recalibrate the contributions of different channel-wise features on each region. The experimental results on three real-world traffic raster flow datasets demonstrate that the predictive performance of the 3D-CANet model was better than the others,which proved the validity of the model proposed.