Traffic Flow Prediction Model Based on Multi-Angle Attention Convolutional Network

Traffic flow prediction is widely used in traffic control, resource scheduling platforms, and intelligent transportation systems (ITS) due to the complexity and dramatic increase in the size of the urban traffic network structure, the continuous growth in the travel demand of the residents, and the increasing traffic congestion problem. Existing traffic flow prediction models divide the city into many equal-sized grids, each representing a region, and use historical traffic data to predict future traffic conditions. However, this approach faces a crucial challenge in extracting complex spatio-temporal correlations within the city, and existing methods often fail to capture the complete spatio-temporal dependencies. To enhance the existing traffic flow prediction model, a new traffic flow prediction model called Multi-Angle Attention Convolutional Neural Network (MACACNN) is proposed. The model extracts spatio-temporal correlations from multiple angles to capture richer information. Specifically, MACACNN consists of three slice-view CNNs: Front Slice CACNN, Side Slice CACNN, and Top Slice CACNN. These slice-view CNNs can extract spatiotemporal correlations between different traffic patterns and establish the spatiotemporal relationship of traffic flow effectively by incorporating a Coordinate Attention mechanism (CA-Net) within the residual units of the branch networks. Furthermore, there is a fusion layer that fuses the results of the three slice view CNNs. In addition, MACACNN its external factors are added to the fusion layer to improve the prediction performance. Experimental studies are conducted on the TaxiBJ public dataset, and the results demonstrate that this approach can improve prediction performance with certain performance gains over other baseline models.