Prediction of urban expressway total traffic accident duration based on multiple linear regression and artificial neural network

Traffic accident duration prediction provides an important basis for traffic mitigation measures after accidents. An accident duration identification method based on velocity thermogram has been established in order to obtain the traffic recovery time after the accident vehicles were removed from the scene. The AIC (Akaike Information Criterion) and the BIC (Bayesian Information Criterion) were applied to fit the probability distribution of the accident duration, and the results showed that the lognormal distribution fitted best. Two multiple linear regression models predicting total duration and clearance time respectively have been constructed based on 9 variables including temporal, spatial, environmental, traffic and accident detail variables. Results showed that traffic condition, location type, accident type and police presence significantly affected total accident duration, while response time and accident type played a significant role in predicting clearance time. After deleting the samples with too short duration, the performance of the two models were both significantly improved, and MAPE were 27.1% and 49.8% respectively. In order to test the performance of the multiple linear regression model, two artificial neural network (ANN) models were also established for comparison. The results of ANN showed more prediction errors and less stability. In general, the multiple linear regression models performed better than ANN.