| Traffic flow prediction is an important part of an urban intelligent transportation system.Efficient prediction algorithms will help people within the city to meet their travel needs,while reducing the waste of urban traffic resources and the probability of traffic congestion.This thesis uses deep learning and Bayesian correlation theory to model the traffic flow dataset,introduces Bayesian deep network and residual calibration deep network methods respectively,and designs two kinds of traffic flow prediction models for spatiotemporal images.Deep mining with spatiotemporal features,using public datasets to evaluate traffic flow prediction performance,and developing a traffic flow prediction system for visual analysis and technical verification.The contributions of this thesis are as follows:First,a traffic flow prediction model(Res Bayes Net)based on Bayesian deep learning is proposed.First,extract the proximity,periodicity and trend of the traffic flow raster space-time image,mark the weather and holiday features,and form a tensor matrix fused with multidimensional features and normalized;secondly,feature splicing of the tensor matrix;The Bayesian staggered deep network is constructed based on the multi-layer Bayes Residual unit;finally,the error distribution is calculated and the weight parameters are updated,and the Res Bayes Net model is established for traffic flow prediction test and performance verification.The experimental results show that the prediction performance of the Res Bayes Net model is better than that of the traditional model,and the root mean square error index of traffic flow prediction is reduced by 0.3 or more.Then,a combined model(MS-Res Cnet)based on residual calibration network and multi-scale fusion mechanism is proposed to realize the prediction method based on spatiotemporal images.Firstly,the time series division method is used to extract the proximity,periodicity and trend of regional traffic flow spatiotemporal images in stages,and CNN is used to obtain the proximity,cycle and trend characteristics of multi-dimensional mosaic matrix;secondly,the spatiotemporal images are extracted respectively through the dual-channel Res Cnet network The benchmark feature and calibration feature of the dataset;thirdly,through the deep staggered training network,the cross-modal spatiotemporal feature extraction is realized,and the spatiotemporal feature matrix based on multi-scale fusion is established;finally,the traffic prediction of the multi-scale feature matrix is performed by CNN,and the public data set is used.Evaluate and validate model performance.The results show that the traffic flow prediction performance of the proposed model is better than that of the three traditional models,and the improved traffic flow prediction root mean square error(RMSE)index is reduced by 0.4,and better prediction results are obtained.Finally,based on the Flask and Echarts frameworks,an urban traffic flow visualization analysis system is built to realize the traffic flow visualization analysis.It supports the import of open source traffic datasets,and provides visual analysis functions such as grid historical traffic flow,predicted traffic flow,boarding and boarding behavior,traffic flow images combined with maps,and prediction performance evaluation charts.Based on the analysis results provided by the system,it provides the technical foundation and platform support for the prediction and analysis of urban traffic flow and travel behavior. |
