Dynamic spatial-temporal graph convolution network for e-bike traffic flow forecasting

Electric bicycles (e-bikes) are playing increasingly important roles in satisfying people's personal mobility and short trips in an eco-friendly (reducing carbon dioxide emission) and efficient way. However, the popularity of e-bikes brings new challenges to public transport management. Fine-grained and accurate forecasting of traffic flow can help provide scientific decision support for urban planners, thereby alleviating many urban management problems, such as traffic congestion, parking lots and charging stations planning, etc. However, traffic flow forecasting is a highly non-linear problem with complex spatial and temporal dependencies. In particular, people usually have more flexible and random mobility when riding e-bikes. Existing methods do not effectively utilize the trajectory data of vehicles or capture the dynamic changes of spatial-temporal dependencies. In this paper, we propose a Dynamic Spatial-Temporal Graph Convolution Network (DSTGCN) for e-bike traffic flow forecasting. Specifically, we have designed an inter-road transition flow embedding module to extract spatial attention matrix between roads, in response to the flexible mobility of e-bikes. To cope with the dense urban road network, we combine graph convolution with the spatial attention matrix, making the model more efficient and refined in capturing spatial-related information. Then, a channel attention component is designed to enhance the ability of the graph convolution module to model the dependencies between channels explicitly. Finally, we conducted extensive experiments on both the real-world e-bike dataset and simulation scenario dataset, demonstrating that DSTGCN outperforms the baselines, including several state-of-the-art methods. Copyright © 2024 IEEE.