Research On Travel Time Prediction Method Of Urban Road Network Based On Deep Learning

Travel time is an important indicator to measure the operation efficiency of urban road traffic system.Accurate prediction of travel time can provide timely and effective road condition information and route planning for travelers,so as to alleviate or avoid traffic congestion.However,on the one hand,the spatial and temporal dependency of actual traffic flow has dynamic time-varying characteristics,and the traditional data modeling methods cannot effectively express the dynamic spatio-temporal correlation of road network traffic,which increases the difficulty of accurate prediction of travel time.On the other hand,affected by external factors such as weather and traffic accidents,the time series data of traffic flow is prone to abnormal fluctuations,which makes it difficult to effectively extract potential key traffic information.As a result,the prediction results are unstable and cannot meet the actual application requirements.Based on the above problems,in this study,the urban traffic multi-source spatio-temporal data is taken as the basis,the deep learning method is adopted as the main mean,and the traffic flow theory and data mining technology are combined to conduct in-depth research on road network travel time prediction.Under normal traffic conditions,the paper aims to solve the problem of modeling the dynamic spatio-temporal correlation of urban traffic.Under abnormal traffic conditions,the paper aims to establish a response algorithm to perceives the abrupt characteristics of traffic through traffic event analysis and modeling,so as to solve the problem of low accuracy and instability of travel time prediction.The specific research contents are as follows:(1)The spatial and temporal characteristics of traffic are analyzed and modeled based on urban multi-source data.Firstly,the road network data,traffic flow data,traffic event data and meteorological data are cleaned,repaired and normalized.Based on this,correlation analysis method is used to choose the optimal time granularity of data aggregation,and a spatio-temporal matching method based on spatial information and time granularity is proposed to solve the problem of spatio-temporal semantic information inconsistency of multi-source data.Then,correlation function and time series decomposition algorithm are utilized to analyze the proximity,daily period and weekly period properties of traffic flow in temporal dimension.Through the Global Moran index and Pearson correlation coefficient,the spatial global correlation and local heterogeneity of traffic flow in each road segment are analyzed.Finally,the fourth-order spatio-temporal tensor model is established based on the hourly mode,daily mode,weekly mode of temporal dimension and road segment mode of spatial dimension.The spatial and temporal features of traffic flow in different modes are extracted by means of fiber and slice of tensor.(2)Considering the impact of traffic events(such as traffic accidents,road construction,etc.)on road travel time,the systematic framework of traffic event detection and duration prediction is constructed.Firstly,a two-stage framework for traffic event detection including binary classification and multi-classification is established.In the first stage,the characteristic variables of traffic flow before and after event as well as upstream and downstream segments are screened from the perspectives of time and space respectively,and the XGBoost binary classification model is built for automatic traffic event detection.In the second stage,aiming at the problem of unbalanced event data categories,three experimental sample sets are established by using random oversampling,weighting method and SMOTE algorithm.Further,XGBoost multi classification algorithm is combined to identify traffic event types.The results show that SMOTE-XGBoost model has the best classification effect.Then,the event duration is accurately predicted by using stacked denoising autoencoder and logistic regression algorithm,and the highest accuracy reached 82.37%.Finally,the systematic framework is integrated into the travel time prediction model,which provides effective characteristic variables for the accurate prediction of travel time.(3)A travel time prediction method based on dynamic spatio-temporal graph convolutional network is proposed to solve the problem of dynamic spatio-temporal correlation modeling of road network traffic under normal conditions.Firstly,a modeling method considering timevarying spatial correlation is proposed.The distance weight matrix and the correlation weight matrix are calculated by using the distance information of road segments and the historical average speed.Next,a hybrid neighborhood matrix is constructed by combining the adjacency matrix of graph,so as to establish an improved graph convolutional network and realize the fusion of spatial static and dynamic characteristics.Then,a classification modeling method of temporal correlation is proposed,which is to extract the temporal characteristics of traffic flow by establishing recent component,daily period component and weekly period component.Based on this,the improved graph convolutional network is applied in each component to construct a prediction framework of dynamic spatio-temporal graph convolutional network.Finally,the model optimization and performance verification are carried out through experiments and case analysis.The highest prediction accuracy of the model is 88.77%,and the goodness of fit of the prediction curve reaches 0.8962.The results show that the proposed method can effectively extract the dynamic spatio-temporal characteristics of road network traffic,and achieve accurate travel time prediction.(4)A travel time prediction method based on multi-component fusion network is proposed to solve the unstable prediction problem under abnormal traffic conditions.Firstly,a hybrid neighborhood matrix with dynamic spatial attention weight is constructed by using the threshold Gaussian kernel function and the dynamic correlation coefficient of road segments.The global attention mechanism is introduced into the Seq2 Seq framework to capture the dynamic dependencies between time series.Based on this,a dynamic spatio-temporal attention component is established in combination with GRU to extract the dynamic spatio-temporal characteristics of traffic.Secondly,the established traffic event detection and duration prediction frameworks are integrated,and a traffic event sensing component is built to extract the characteristics of traffic events.Thus,a response prediction algorithm based on "gating" fusion is developed.Through gating multi-source fusion components,the abrupt characteristics of traffic are captured and multi-source information fusion is realized.In single-step and multistep prediction,the model accuracy is improved by 2.01% and 2.92% respectively.Finally,the function of each component is illustrated by ablation experiment.Through comparative experiments and case analysis,it is verified that the proposed model can achieve accurate and stable travel time prediction under abnormal conditions.This paper takes the road network around Beijing Olympic Sports Center as the research object,and focuses on the dynamic spatio-temporal correlation modeling of traffic and the unstable prediction of travel time under abnormal conditions.The research results can help improve the travel prediction ability of urban traffic system under different traffic conditions,which has positive significance for increasing traffic operation efficiency,improving traffic safety,and promoting green and low-carbon cycle development.