Research And Application Of Traffic Flow Prediction Method Based On Deep Learning

With the year-on-year growth of China's car ownership,traffic congestion has become an inevitable problem in the process of urban development.Research shows that the key to managing congestion is not to build more roads,but how to rationally use limited transportation resources to improve the efficiency of urban road networks.Therefore,reasonable roadworks and traffic signal timing have become the primary tasks of traffic management departments,and the realization of the above can not be separated from the traffic flow prediction technology.By predicting the state of traffic flow and sensing the roads that will be congested in the road network,traffic flow prediction technology can provide powerful data support for congestion management.There are many types of algorithms and models for traffic flow prediction,but the complex road structure of urban road networks and the highly nonlinear nature of the traffic system make it difficult for traditional prediction methods to be applied on a large scale in cities.The data-driven deep learning method can better solve the problems of timeliness and accuracy brought by the large amount of data processing,effectively extract the complex features of the data,and improve the reliability of traffic flow prediction.In view of this,dissertation conducts research on traffic flow prediction methods based on deep learning.First,in order to make full use of traffic data from multiple sources,a traffic data fusion method based on virtual road network and virtual time series is proposed to unify the spatiotemporal attributes of multi-source traffic data to obtain more comprehensive traffic flow information.The data set serves as the data basis for subsequent model research.Secondly,in order to accurately identify traffic congestion,the characteristics and applicable scope of commonly used traffic state evaluation indicators are analyzed,and on this basis,the Traffic Stress Index(TSI)based on road V/C and the Delay Index(Delay)based on travel time ratio are constructed.Aiming at the fuzzy nature of traffic flow state division,fuzzy mathematical method was used to make a fuzzy comprehensive evaluation of traffic flow state based on the above two types of indexes.Finally,aiming at the problems of slow training and over-fitting of ordinary LSTM deep learning models when dealing with large-scale data,the attention network was introduced to build a DA-LSTM short-term prediction model of traffic flow with a two-stage attention layer.,and based on the fused traffic data,the model's prediction effect is tested.Result shows that the model can achieve simultaneous prediction of multi-lane and multiple traffic flow parameters,and the prediction effect is better than ordinary LSTM and GRU models.Based on the above research work,combined with user needs,a traffic congestion warning system was designed to alert intersections in the urban road network where congestion has occurred and may occur in the form of intersection congestion alarms,helping traffic signal timing personnel to quickly locate and control Intersections with unreasonable plans can more effectively avoid the spread of traffic congestion.