Research On Traffic Prediction Technology Based On Spatial-temporal Convolution Neural Network

With the continuous acceleration of urbanization,the increasing population and the significant mobility of mobile devices have caused a series of problems such as traffic congestion and resource allocation difficulties,posing severe challenges to intelligent urbanization systems.As an important component of intelligent control systems,traffic prediction technology aims to predict future traffic information using observed historical data.Managers can make reasonable traffic scheduling based on the prediction results and respond to possible emergencies in a timely manner to improve management efficiency and reduce resource waste.However,achieving accurate traffic prediction is challenging,as existing models have difficulty comprehensively mining the temporal and spatial characteristics of traffic data and maintaining high prediction accuracy in long-term forecasting.In addition,existing research cannot handle the problem of data differences between various regions in multiregion collaborative traffic prediction effectively.This paper proposes three effective spatial-temporal traffic prediction frameworks to address the challenges faced by existing models:1.An adaptive spatial-temporal convolution network model is proposed for traffic prediction,which mainly consists of two components:adaptive temporal convolution and adaptive spatial convolution.The adaptive temporal convolution network extracts historical traffic information using an adaptive time transition matrix,assigning different feature aggregation parameters to different timestamps to reflect their heterogeneity,thereby achieving both long-term and short-term prediction accuracy improvement.Adaptive spatial convolution can adaptively extract global spatial correlations based on specific prediction tasks without any prior knowledge.2.An adaptive graph cross-step convolution network model is proposed for traffic prediction,consisting of cross-step convolution and dual-channel adaptive spatial convolution.Cross-step convolution extracts time features through multiple parallel cross-step convolution kernels and assigns different convolution kernels to reflect temporal differences in prediction timestamps,thus achieving both long-term and short-term prediction accuracy improvement.Dual-channel adaptive spatial convolution further enhances the model’s ability to extract spatial features by simultaneously constructing trend and periodicity parts of spatial correlations based on adaptive spatial convolution.3.A distributed multi-region collaborative traffic prediction framework based on federated transfer learning is proposed.Based on the use of recurrent neural networks and graph convolution neural networks to extract time and spatial features,federated transfer learning is used to aggregate global models in high-quality regions and transfer the trained global models to low-quality regions,achieving effective collaborative traffic prediction in each region.Experimental results on multiple public datasets show that the traffic prediction accuracy of adaptive spatial-temporal convolution and adaptive graph cross-step convolution is significantly better than existing baseline models,achieving optimal prediction accuracy for all timestamps within the prediction range.In addition,in the multi-region collaborative prediction framework,federated transfer learning achieves overall prediction optimization in each region without leaking data privacy.