Expressway Traffic Incident Detection And Duration Prediction Based On Incomplete Information

With the rapid development of social economy,traffic demand has increased dramatically,and the problems of traffic congestion and road traffic safety are increasing.Strengthening the construction of intelligent transportation system has become one of the effective ways to solve traffic problems.Specifically,the perception and prediction of traffic flow state is an important support for the intelligent and informational travel service of intelligent transportation system,and the automatic detection of traffic incident and the prediction of incident duration have become the key parts of traffic flow state perception and prediction.Under the background of traffic big data,how to make full use of the massive traffic information data of various traffic detectors and mine the hidden effective information has become the research focus of automatic traffic incident detection and incident duration prediction.At the same time,in practical research,these traffic big data information are often incomplete.For example,traffic incident data due to the accidental nature of the incident,the amount of data is scarce and unbalanced;the acquisition time of traffic incident variables is different,and the variables cannot be obtained in the first time.Based on this incomplete information environment,this study uses machine learning as a tool and various traffic data as data support,combined with the spatial and temporal characteristics of traffic flow and data mining technology to conduct in-depth research on traffic incident automatic detection and incident duration prediction.It aims to solve the problem of small and imbalance sample data in incident detection and the lack of variables for duration prediction.The specific research contents are as follows:(1)Expressway traffic incident detection based on ensemble learning under imbalanced and small datasetsFirstly,based on the spatial and temporal characteristics and real-time performance of traffic incident,57 variables are selected to form the initial variable set;Then,the SASYNO oversampling technique is used to balance the data set for the initial variable set;random forest algorithm extracts feature variables;finally,the ensemble learner RSKNN is trained for classification detection with feature variables as input.(2)Prediction of duration of traffic incident by hybrid deep learning based on multi-source incomplete datasetsFirstly,based on incomplete data environment,this study combines natural language data to improve the prediction accuracy of the model.The LDA topic model is used to extract variables from natural language data;the extracted variables are input into the Bi-LSTM network model,and feature fusion is performed with the LSTM network through the connection layer;finally,the full connection layer is input to the regression layer for prediction.In addition,considering the difference of data acquisition time,a staged sequential prediction model is established by hybrid model and K-means clustering algorithm to meet the requirements of real-time prediction.(3)Empirical research of incident detection algorithm and duration prediction modelThe SASYNO-RF-RSKNN incident detection algorithm is experimentally analyzed by two real road data sets,I-205 and I-880.Firstly,the data set is balanced and standardized by SASYNO algorithm.The variables after RF feature selection are classified and detected.Eight machine learning algorithms are compared on six performance evaluation indexes.Whereafter,the proposed deep hybrid duration prediction model is verified by a public large-scale data set.Using RMSE,MAE and MAPE as error evaluation indicators,seven machine learning prediction algorithms are compared,and the number of LDA topic models is also tested.Finally,the staged sequential prediction model based on hybrid network is verified and analyzed.