Traffic Flow Prediction And Congestion Prevention Via Deep Learning

With the fast development of city and its road networks,the number of motor vehicles and road traffic are rapidly increasing accordingly,and hence traffic congestion has become one of the critical problems in urban development and management.Early solutions to traffic congestion are based on the real-time road traffic flow information,which often lag behind and hence cannot imporve the road conditions timely.Therefore,it is necessary to forecast the road traffic flow to make a timely congestion prevention plan.However,the data structure of traffic flow is often complex and the data volume of traffic flow is often huge.Moreover,the traffic flow data has a strong spatio-temporal correlation.Due to these factors,it is very difficult to extract the valid information about future traffic flow from the huge existing traffic flow data.In recent years,many researches have shown that deep learning is a promising method to process a large amount of data.Thus,the focuses of this thesis is study how to make use of the existing road traffic flow information to predict the traffic flow in the near future by means of deep learning,and how to use the predicted traffic flow information to make effective traffic congestion prevention solutions.The main purpose of traffic flow prediction is to predict the traffic flow information of a particular section in the near future,according to the current and historical traffic flow informaiton observed at different locations of the road network.We analyze several basic models of deep learning,and choose the deep belief network as the basic model of traffic flow prediction,which has the advantages of fast training and good fitting of complex functions.The training of the model requires some sample data,but the real traffic flow data sample is often hard to obtain.To solve this problem,we build a simulative road network in Shenzhen using the traffic simulation platform,and obtain the simulative traffic flow data for learning and training.Moreover,we adopt the divergence algorithm to train the deep belief network.The goal of training is to maximize the fitting of traffic flow data.In the training process,we need to adjust the following four important parameters: the number of nodes in the input layer,the number of hidden nodes,the number of nodes in the output layer,and the number of network layers.In this thesis,we provide the detailed method for the adjustment of these parameters and obtain the optimal belief network structure for the traffic flow prediction.We further compare our proposed method with the existing methods,and show that our proposed method based on deep learning can achieve a much better performance in term of the accuracy of the prediction.According to the predicted traffic flow,we further propose two traffic congestion prevention solutions.The first one is to adjust the time lengths of associated traffic lights,so as to slower the traffic flowing into the congested roads and speed up the traffic flowing out of the congested roads.The second one is to set congestion fees on the congested roads,so as to reduce the volume of traffic flowing into the congested roads.Simulation results show that the proposed solutions can not only reduce the lane occupancy rate of the congested road sections,but also reduce the average lane occupancy rate of the entire road network,hence increasing the traffic volume of the whole road network and optimizing the road network resource utilization.