Traffic State Recognition And Prediction Based On Vehicle Trajectory Data

With the rapid increase of motor vehicle ownership,the imbalance between traffic demand and supply and the problem of traffic congestion become more severe.How to efficiently utilize road capacity,mitigate traffic congestion and ensure smooth,efficient,green and safe operation of urban traffic is one of the most important topics in traffic management and control,for which the information of the prevailing traffic and the future traffic states are considered as the prerequisites.Therefore,this thesis proposes a traffic state identification and prediction approach based on sample vehicle trajectories provided by the car hailing company Di Di.The main contents of the thesis are as follows.First of all,the raw vehicle GPS trajectory data is processed.According to the data characteristics of the trajectory data,data cleaning and coordinate transformation are conducted;The Arc GIS software platform is used to do topology construction processing on the road network data,and the hidden Markov map matching algorithm is used for map matching to complete the correction and reconstruction of the trajectory data;Based on this,the instantaneous vehicle is calculated,and the statistical period is divided to extract the average travel speed of the road section,the speed variance and the number of vehicle samples,etc.The data is used to construct a new dataset,which provides the data basis for the later research.Then,a traffic state identification approach is developed based on the fuzzy C-means clustering method.On the basis of analyzing and summarizing the related research on traffic state recognition,the average travel speed,the speed variance and the number of vehicle samples of the road section are selected as the characteristic parameters according to the characteristics of the trajectory data.Based on the uncertainty and fuzziness characteristics of the traffic data,the traffic state recognition model is established by applying the fuzzy C-mean(FCM)clustering algorithm.Furthermore,the optimal number of states that meet the actual situation of road traffic in China are determined by referring the relevant information.We use38 days’ field data to train the model to classify traffic states into four categories: free flow,near free flow,slow-moving and congested states.Finally,a traffic state prediction method is proposed.Based on the understanding of existing traffic prediction methods and neural network models,the traffic state prediction model based on LSTM-SVR is proposed,and then the prediction range of the model is determined based on the analysis of the temporal correlation between the average travel speed of the road section,the number of vehicle samples and other characteristic parameters between weekdays and weekends.The optimal parameters of the model are determined by reviewing the literature and pre-training for several times.We choose the data of weekdays to validate the effectiveness of the proposed prediction model.The results show that the proposed LSTMSVR model performs better than the LSTM and the SVR model alone.Moreover,a total of715 groups of states in the test set are analyzed according to the predicted traffic parameters,and 649 groups are successfully predicted,with a successful prediction rate of 90.34%,which indicates the accuracy and effectiveness of the model.