Traffic flow prediction based on graph wave adaptive spatiotemporal graph convolution network

Graph-based traffic flow predictioning is widely applied in traffic systems, where constructing intricate spatiotemporal correlation models from relevant time series data is imperative for comprehending the dynamics of the traffic system. The extraction of features from graphical data, coupled with the integration of time series data, serves to enhance the accuracy of traffic flow predictions. Additionally, the challenge arises when real traffic flow data often contains missing values. Predicting traffic in scenarios with missing data proves to be challenging, as existing traffic flow predictioning methods frequently lack the capacity to model dynamic spatiotemporal correlations in the presence of such gaps, leading to unsatisfactory prediction outcomes. This paper introduces the Adaptive Spatiotemporal Graph WaveNet-based Graph Convolutional Network (AST-GW-GCN) to tackle traffic flow predictioning. AST-GW-GCN comprises three independent components, each modeling short-term, daily, and weekly dependencies of traffic flow. Within each component, Gated Temporal Convolutional Network (TCN) and Graph Convolutional Network (GCN) serve as encoders, conducting spatial convolution to extract spatial correlations and temporal convolution to capture temporal correlations, thereby generating hidden features. The Gated Recurrent Unit (GRU) is employed to decode these hidden features and weigh the outputs of the three components, producing the final prediction result. The spatial convolution module establishes an adaptive adjacency matrix to overcome the physical constraints of the graph structure, facilitating improved extraction of hidden spatial dependencies within the data. Furthermore, experiments are conducted under various data missing patterns and missing ratios. The experimental findings, based on the PeMS08 real dataset, demonstrate that the proposed AST-GWGCN comprehensively captures spatiotemporal correlations in the data, outperforming baseline models in terms of performance.