Research On Identification And Prediction Of High-risk Urban Crash Spots Based On The Spatio-temporal Dimension

With the increase in urbanization,motorization,and population in urban regions,the urban road traffic safety situation has become severe and complex,which has brought tremendous pressure to traffic management departments.Urban motor vehicle crashes cause loss of life and property and induce or aggravate urban road network congestion,resulting in potential negative impacts on society.Identifying high-risk urban crash spots and implementing targeted countermeasures are the most cost-effective method to improve the overall traffic safety level of urban road networks.However,the existing researches still mainly use mathematical statistics to identify high-risk spots,ignoring the inherent spatial and temporal correlations of high-risk spots in urban road networks.High-risk crash spots have class imbalance and multi-factor influence characteristics.Using traditional statistics models easily leads to low prediction accuracy and high computational cost,resulting in an insufficient understanding of the causes of high-risk urban crash spots.Therefore,the accurate identification of high-risk urban crash spots,selection of key risk factors,and risk prediction have important research values.They are beneficial in providing a decision-making basis for controlling and preventing high-risk urban crash spots.By integrating spatio-temporal modeling theory and machine learning techniques,this study takes the space-time cube as the basic spatio-temporal unit for analyzing the urban road network.Research has been carried out on identifying highrisk urban crash spots,selecting key risk factors for high-risk urban crash spots,and predicting high-risk urban crash spots.The main research contents of this study include the following six aspects:First,a literature review of the three aspects,including identification of highrisk crash spots,an analysis of urban crash risk factors,and crash risk prediction research,is carried out.Urban crash risk sources and risk factors are reorganized,the advantages and disadvantages of crash risk prediction models are summarized,and popular trends and existing bottlenecks in the related research are analyzed.Second,the space-time cube model in spatio-temporal modeling theory is introduced.Definition,classification,and quantification methods of crash risk are summarized,and the calculation method of crash risk based on space-time cubes is proposed.Crash risk and risk factor characteristics are mainly analyzed,and the candidate features of urban crash risk based on space-time cubes are obtained based on the current literature.Third,the principle and steps of the emerging hot spot analysis method based on spatial statistics theory are illustrated.Limitations of the emerging hot spot analysis method are discussed.An improved emerging hotspot analysis method based on machine learning is proposed to solve the false positive problem of using local spatiotemporal statistical indicators to identify hotspots.The accuracy of identifying highrisk urban crash spots in the dimensions of space and time is improved by designing machine learning steps such as candidate feature extraction,risk label generation,and random forest classifier construction.Fourth,the framework for selecting key risk factors of high-risk urban crash spots based on space-time cube-machine learning(STC-ML)is constructed to overcome the class imbalance problem.The framework illustrates the selection of key features,crash risk pattern prediction,and key feature explanation.In this process,the resampling technique,embedded-based feature selection method,classifier training,and parameter tuning are discussed.Mathematics of the addictive feature attribute method and SHAP(Shapley Additive ex Plannations)method of interpretable machine learning are analyzed.Fifth,the method for predicting high-risk urban crash spots based on time-series deep learning is proposed.Time-series clustering techniques,multivariate time series dataset construction based on sliding window,and multi-layer long short-term memory neural network prediction model are discussed.A Bayesian optimization method is used to optimize the structure and hyperparameters of the time-series deep learning prediction model,and the prediction performance of the deep learning model after Bayesian optimization is evaluated.Sixth,Manhattan in the United States and Shenzhen in China are typical case cities at home and abroad to verify the feasibility and practicability of the data-driven methodology in this study.Basic data related to urban crashes in the two case cities are collected.Identification of high-risk urban crash spots,selection of key risk factors,and prediction of high-risk urban crash spots are carried out.Specific control and prevention suggestions for high-risk urban crash spots in case cities are put forward,and the similarities and differences of case cities are compared.The results of this study show that: using the improved emerging hot spot analysis method based on machine learning can avoid the problem of false positives in large samples and effectively improve the identification accuracy of high-risk urban crash spots in the spatio-temporal dimension;key features of traffic facilities,public transit network,land use,and traffic flows have the greatest contribution to high-risk urban crash spots,and the key features of spatial neighborhoods account for the majority of selected key features,indicating that spatial neighborhood features have important impacts on hot spot patterns;the smaller spatial step of space-time cubes,the more obvious and complex the interaction effect of key features on hot spot patterns;the method based on time-series deep learning can identify high-risk clusters with an upward trend in the future,effectively solve the problem of trend,seasonality,periodicity,and instability in the time series,and accurately predict the risks of highrisk urban crash spots.In addition,typical case cities at home and abroad have further verified the feasibility and practicability of the data-driven methodology framework of this study,providing new ideas and new insights into machine learning for urban traffic safety management.It helps to improve the traffic safety management capabilities and the level of data governance of urban traffic management departments,thus effectively reducing the negative impact of urban crashes on the whole society.