Research On Traffic Condition Prediction Based On Spatiotemporal Convolution Mode

The rapid rise in private car ownership has led to increasing traffic congestion.By making real-time predictions of traffic conditions in the future period,decision-making guidance can be provided for the timely evacuation of traffic flow,thereby effectively reducing the occurrence of traffic congestion.However,due to the serious spatial-temporal dependence and strong real-time requirements of traffic prediction.Therefore,the prediction of traffic conditions not only has higher requirements for the accuracy of the model but also has higher requirements for the running speed of the model.The existing model cannot balance the speed and accuracy,and cannot take into account the two requirements well.Therefore,existing models are still difficult to adapt to the practical application of large-scale road networks.This makes traffic condition prediction still challenging.Based on the above requirements,this thesis proposes solutions for traffic condition prediction in two processing dimensions: on the one hand,this thesis improve accuracy and operational efficiency of the model by innovation in the design of the model;On the other hand,by learning the relationship between traffic congestion types and traffic speeds of road,this thesis directly classifies the traffic congestion situation in the future period and realizes a more intuitive prediction of traffic conditions.The main work of this thesis includes the following points:(1)In this thesis,a dual-channel temporal convolutional layer and a spatial gated convolutional block are designed to extract spatial and temporal features respectively.Spatial gated convolutional blocks can control the amounts of spatial features and retained original temporal features through gated mechanisms.This allows the model to retain the original temporal features during spatial feature extraction.So that subsequent dual-channel temporal convolutional layers can acquire more efficient temporal features.By stacking multiple dual-channel residual temporal convolution blocks with different dilated rates to compute spatial-temporal features of different dimensions,the prediction accuracy of the model can be effectively improved.The design of the full convolutional structure makes the model more sensitive to some dynamic changes so that the model can timely predict some violent dynamic changes.Multiple comparative experiments on real-world datasets have demonstrated that the model presented in this thesis can predict the speed of the future more quickly,accurately,and stably.(2)In this thesis,a deep learning model for the simultaneous prediction of dual traffic metrics is proposed.Leverage the correlation between the two metrics to provide each other with more effective temporal and spatial features.Through the effective extraction and flexible integration of spatial-temporal features,the prediction effect of two traffic metrics can be improved at the same time.Based on the dual-channel residual temporal convolution block,we designed the dual-channel gated convolutional block.Moreover,when inter-layer feature maps are fused,the inter-layer attention mechanism is added.This allows for more efficient extraction and more adaptive fusion of different dimensional features.(3)This thesis proposes a new spatiotemporal convolutional model to learn the nonlinear mapping relationship between traffic congestion category and speed information,and the mapping relationship is applied to predict the traffic congestion category in the future.The model changes the previous processing method of traffic congestion category prediction through threshold setting.Moreover,in order to solve the problem of unbalanced number of samples in each category in the traffic congestion category data-set,this thesis designs data-sets with different category sample proportions to train base learners with different sensitivity to different categories,and cooperates with the weight update algorithm of base learners to make effective use of the advantages of each basic learner,so as to achieve a more stable and accurate prediction effect.On two real-world data sets,the speed prediction model SDTCN proposed in this thesis compare with the best-performing speed prediction model GMAN,the MAPE is reduced by 5.35% on average.At the same time,the training time of the SDTCN is reduced by 22.5% compare to STGCN.In order to further improve the prediction accuracy of the model for speed prediction,this paper proposes a deep learning model SEDTCN for simultaneously forecasting dual traffic metrics.Compared with the SEDTCN model that predicts a single traffic metric,the MAPE values corresponding to the TTI and Speed traffic metrics implemented by the SEDTCN model that predicts dual traffic metrics are reduced by 35.1% and 15.3% respectively.Based on the accurate prediction of speed,this paper proposes an integrated learning model that learns the nonlinear mapping relationship between road traffic speed and traffic congestion categories to achieve more intuitive traffic condition prediction.Compared with the SE-DCGCN single model on the two traffic congestion data sets of Shenzhen and Xi’an city by the F1-Score corresponding to congestion category prediction,the ensemble model of SE-DCGCN has increased by 34%.