Real-time regional risk prediction can play a crucial role in preventing traffic accidents. Thus, this study established a lane-level real-time regional risk prediction model. Based on observed data, the least squares-support vector machines (LS-SVM) algorithm was used to identify each lane region of the mainline, and the initial traffic parameters and surrogate safety measures (SSMs) were extracted and aggregated. The negative samples that characterized normal traffic and the positive samples that characterized regional risk were identified. Mutual information (MI) was used to determine the information gain of various feature variables in the samples, and the key feature variables affecting the regional conditions were tested and screened by means of binary logit regression analysis. Upon screening the variables and corresponding labels, the construction and verification of a lane-level regional risk prediction model was completed using the catastrophe theory. The results showed that lane difference is an important parameter to reduce the uncertainty of regional risk, and its odds ratio (OR) was 16.30 at the 95% confidence level. The 10%-quantile modified time to collision (MTTC) inverse, the speed difference between lanes, and 10%-quantile headway (DHW) had an obvious influence on regional status. The model achieved an overall accuracy of 86.50%, predicting 84.78% of regional risks with a false positive rate of 13.37% and 86.63% of normal traffic with a false positive rate of 15.22%. The proposed model can provide a basis for formulating individualized active traffic control strategies for different lanes.
Keywords: catastrophe theory; feature analysis; regional risk prediction; roadside observation data; surrogate safety measure.