Deep Urban Traffic Prediction Research Based On Spatio-Temporal Data

With the advancement of urbanization,intelligent transportation systems（ITS）facilitate urban residents to travel and promote economic and cultural development of cities.Accurate traffic forecasting is essential for building an intelligent transportation system,providing suggestions for traffic planning,residents’travel,and operations departments to make decisions.The field of traffic forecasting includes the prediction of travel demand,traffic flow and traffic speed.This thesis investigates travel demand and traffic flow forecasting with the following main contributions:For travel demand prediction,this thesis proposes a travel demand forecasting model based on Markov cluster algorithm and spatio-temporal graph attention network（MCL-STGAT）.The model consists of three modules:a temporal block,a spatial block,and a prediction layer.In spatial block,this thesis proposes a new method to construct a traffic semantic correlation graph based on Markov cluster algorithm.The semantic correlation matrix is constructed based on the traffic semantic correlation graph to capture deep semantic information of traffic data.Considering complexity of traffic data and nature of non-Euclidean structure after transformation into graph data,this thesis uses a graph attention layer based on the Node2Vec graph embedding algorithm and a convolutional layer based on the Markov cluster algorithm to obtain deep spatial dependencies.A long-short time memory network is used in time block to capture temporal dependencies of traffic data.The performance evaluation of travel demand forecasting model proposed in this study on the New York City yellow cab dataset shows that mean absolute error（MAE）is 0.35 and root mean square error（RMSE）is0.45.The performance evaluation of travel demand forecasting model proposed in this study on the Chengdu city online car dataset shows that MAE is 0.028 and RMSE is 0.048.Experimental results show that the MCL-STGAT is more accurate in prediction than other baseline models.For traffic flow prediction,this thesis proposes an urban traffic flow forecasting model based on generative adversarial network（TGAN）.The model consists of two parts:a spatial feature extraction module and a generative adversarial network framework.A convolutional layer based on the Node2Vec graph embedding algorithm is used in spatial feature extraction module to capture deep spatial dependencies.A new method is proposed to construct traffic embedding graph based on the Node2Vec graph embedding algorithm,and the obtained embedding matrix makes the traffic data suitable for convolutional neural network processing.The generative adversarial network framework is used to capture underlying patterns of how traffic flows evolve in response to changes in travel demand.This study uses the New York City yellow cab dataset and the Chengdu city online car dataset for experiments.The evaluation metrics:Euclidean distance between real and generated flow distributions（D₁）,MAE and RMSE are used to measure difference between generated traffic distribution and real traffic distribution.D₁,MAE and RMSE obtained from the New York City yellow cab dataset are425.73,4.86 and 8.71,respectively.D₁,MAE and RMSE obtained from the Chengdu city online car dataset are 597.47,9.56 and 16.59,respectively.The experimental results show that the TGAN predicts more accurately than baseline models.