Spatiotemporal Adaptive Hybrid Graph Convolutional Networks for Traffic Flow Forecasting

Abstract

Traffic flow prediction is crucial for intelligent transportation systems to reduce congestion. However, accurately forecasting traffic flow is challenging because of its complex spatial and temporal relationships. Existing methods often fail to handle non-linear spatial-temporal dependencies and struggle with large-scale data processing. To overcome these issues, this paper proposes a new spatiotemporal adaptive hybrid graph convolutional network (STAHGCN) for traffic flow prediction. The STAHGCN consists of two main components: an adaptive hybrid graph convolutional module (AHGCM) and a gated temporal convolutional network (TCN). The gated TCN uses dilated causal convolutional networks at different granularities to capture temporal dependencies. AHGCM integrates spatial gate fusion mechanisms, dynamic graph learning (DGL), and static adaptive graph learning (SAGL) to capture dynamic spatiotemporal features of historical traffic flow. DGL reduces time complexity through parallel processing and captures hidden spatial correlations using multi-head graph attention mechanisms. SAGL adaptively captures dynamic spatial properties of traffic. Experiments show that STAHGCN outperforms baseline models. In PEMS-BAY, STAHGCN improves mean absolute error and root mean squared error by 15.49% and 9.88%, respectively, compared with the state-of-the-art model.