Real-time Forecast Models for Traffic Accidents on Expressways Using Stability Coefficients of Traffic Flow

Real-time forecast models for traffic accidents requires a large number of variables, which causes difficulties in data collection and decreases reliability of the model due to overfitting. Two interpretable variables, vertical and horizontal stability coefficients of traffic flow, are proposed to simplify the set of variables, which can facilitate the implementation of forecast models for traffic accidents and reduce the effects of overfitting. Three algorithms including support vector machine, random forest, and logistic regression are selected to develop real-time forecast models for traffic accidents on expressways, respectively. The experiments are conducted based on data of traffic accidents and historical traffic flow collected from the expressway G3001 in the city of Xi'an. In addition, the improved GI index is used to evaluate the significance of the proposed two stability coefficients of traffic flow. The effects of the two proposed coefficients on reducing overfitting is verified through comparing accuracies and AUC values of the set of variables in the test and training data.Besides, the computational efficiency is evaluated by the training time to verify the reliability of the developed models with the two coefficients. The results show that the improved GI indices of the models with horizontal and vertical stability coefficients of traffic flow are 0.952 and 0.922, respectively, which indicates that the proposed two coefficients are more significant for forecasting accidents on expressways than other variables. In the three models, the simplified set of variables based on the two coefficientshas prediction accuracy of 91.1% and 90.5%, respectively, in training and test data, which is similar to the original set of variables. The differences of average prediction accuracy between the simplified set of variables and the original set of variables are 0.69% and 4.87%, respectively. The difference of average AUC values between the two sets of variables are 1.61% and 5.87%, respectively. The average time cost of model training with the simplified set of variables decreases by 15.2%. Thus, the two proposed stability coefficients of traffic flow can improve both the reliability and the computational efficiency of the models.