Traffic Flow Prediction Based On Deep Learning

With the development of society,traffic problems such as traffic congestion have become increasingly prominent.As one of the effective means to solve traffic problems,intelligent transportation systems have become more and more popular.Noteworthily,traffic flow forecasting is the basis of intelligent transportation system.It makes accurate predictions on traffic flow using historical data to provide appropriate advice for traffic planning and passenger travel.However,there are still many shortcomings in current methods traffic flow prediction.Most of the traffic flow prediction methods are mainly based on the time series characteristics,which will result in the lack of information and reduce the accuracy of the prediction.Other prediction methods comprehensively utilize the spatio-temporal feature of a specific location,but the spatial feature extraction method is not perfect,which means the accuracy of the extracted information needs to be improved,and this insufficiency leads to the low prediction ability of the method.Aiming at solving the forementioned problems,this thesis proposes a traffic flow prediction method based on deep learning.After the processing and shaping of data,the graph convolutional network and the long short-term memory network are used to extract the spatial and temporal characteristics of traffic flow,respectively,and then an initial prediction result is obtained,integrating which with two versions of predicting results using time-period and weather information,a final result which is more accurate is achieved.The main contributions of this thesis are as follows:(1)To get over the shortcomings of traditional spatial feature extraction methods for traditional traffic flow prediction,a spatial feature extraction method which is more practical is proposed and it manifests better conformability to the actual traffic network structure.The method is characterized in that it regards the road network as a topological map,in which the traffic flow on the road is adopted as the feature of the edge,whereas the obtained feature turns to the node feature with the application of line-graph theory illustrated in graph theory,through which the road network topology can be transformed into a road adjacency topology,and then the road adjacency topology matrix and the node feature matrix are sent into the graph convolution network to realize a reasonable extraction and utilization of the traffic flow spatial characteristics.The experimental evaluation results based on the taxi vehicle dataset in Beijing show that using the proposed method,the traffic flow prediction accuracy is significantly improved compared with the method based on CNN+LSTM,and the prediction error is reduced by 56.9%.(2)Combining the spatio-temporal features extracted with deep learning method with other related features,a novel traffic flow prediction system based on deep learning is proposed.The system predicts the traffic by reasonably extracting the spatio-temporal characteristics of traffic flow,and then integrates the prediction results with other characteristics related to traffic flow(periodic characteristics,weather characteristics)to obtain the final traffic flow prediction results.The experimental evaluation results based on the taxi vehicle dataset in Beijing show that the traffic flow prediction system proposed in this thesis achieves a significantly-improved predicting accuracy compared with previous method.For instance,compared with the prediction method based on CNN+LSTM,the prediction error in our system is reduced by 62.8%,and with using the method illustrated in(1),the prediction error of our system is reduced by 13.7%.The flow prediction method proposed in this thesis is universal and can be applied in similar spatio-temporal sequence prediction problems on traffic network and information network,so it has certain theoretical research value;in addition,the traffic prediction method and system proposed in this thesis can accurately predict road traffic flow,which means a lot from the perspective of reality and practical application.