Road Network Traffic Forecasting Based On Spatio-Temporal Graph Neural Network

Traffic forecasting is a classic direction and research hotspot in intelligent transporta-tion system,which is helpful for traffic control,route planning,vehicle scheduling and other tasks.It plays an important role in alleviating traffic congestion and ensuring pub-lic transportation safety.The main factors affecting traffic prediction include the spatial correlation of road network structure and the temporal dependence of nonlinear dynamic changes of road conditions.Recently,various deep learning methods have been applied to traffic prediction,mainly employing graph neural networks for spatial correlation mod-eling and recurrent neural networks for modeling temporal dependence.However,most existing methods assume that spatial correlation is static,and the temporal correlation only has sequential dependence,and the uncertainty in the traffic is not fully considered.Aiming at the challenges of existing traffic forecasting,this thesis designs two mod-els,namely,reinforced spatio-temporal attentive graph neural network(RSTAG)and vari-ational graph recurrent attentive neural network(VGRAN),to improve the performance of traffic prediction and quantify the uncertainty in traffic domain respectively.The main research contents and innovative contributions are as follows:(1)Aiming at the complicated spatio-temporal correlation in road network,this the-sis constructs two attention mechanisms to capture the importance of different spatio-temporal factors in traffic by extending graph neural network.Specifically,this method adopts spatio attention mechanism to model the spatial correlation of different locations in the graph network,and utilizes temporal attention mechanism to model the periodic de-pendence of different times in multi-step prediction,which makes the model pay attention to the most important spatio-temporal factors of traffic.(2)Aiming at the problem that current traffic prediction models will accumulate er-rors along with the generation of the time series,this thesis utilizes a reinforcement learn-ing strategy optimization method to update the parameters of our model.Specifically,this algorithm combines the maximum likelihood estimation with the reward function of rein-forcement learning,and gradually improves the prediction performance of the model by using its own generated value in the training process,which alleviates the biased prediction problem of the model and improving the performance of traffic prediction.(3)Aiming at the influence of unpredictable random events in traffic,this thesis studies a variational bayesian technique to quantify the uncertainty of traffic prediction.Specifically,this method applies the variational graph neural network to model the un-certainty of node representation and sensor attributes,and exploits the generative flow model to output the traffic sequence with uncertainty,which overcomes the unreliability caused by the traditional point value prediction method.In addition,it can realize effec-tive variational inference and reliable modeling of implicit posterior of traffic data,which can promote the interpretability of the model.In order to verify the validity and reliability of the proposed model,this thesis makes a full comparative experiment and ablation analysis on two real traffic data sets,and the experiment proves that the method implemented in this thesis are superior to other bench-mark models in performance evaluation.In addition,the research in this thesis not only improves the performance of traffic prediction from the perspective of algorithm opti-mization,but also quantifies the uncertainty of traffic prediction from the perspective of model reliability.Therefore,the method proposed in this thesis has certain universality and continuity,and has essential theoretical research value and practical application sig-nificance.