| Objective In recent years,the prediction of the severity of traffic accidents has attracted increasing attention from researchers and law enforcement agencies.In order to simulate the relationship between road severity results and meteorological factors,many models have been proposed.The purpose of this study is to use random forest and neural network two machine learning models to investigate the impact of meteorological variables on the severity of road traffic accidents,and to compare the performance differences between the two models.Methods The traffic accident data of the traffic police detachment of Suzhou Public Security Bureau and the real-time weather data of the National Meteorological Information Center(http://data.cma.cn)were used.The research period is from January 2008 to April 2017.A total of 7 795 traffic accidents were included in this study.We try to use a random forest model to fit the nonlinear relationship between meteorological variables and the severity of traffic accidents,and compare the prediction accuracy of neural network models.The model is built by the random Forest package and Neuronet package in R software.75% of the training samples are separated from the data to build a predictive model,and the remaining 25% of the test samples are used for testing.In addition,in order to understand the accuracy of the model predictions,the prediction results are calculated and compared with the actual results.Results From 2008 to 2017,the number of motor vehicle owners in Suzhou increased from 242 per 10 000 people to 673 per 10 000 people.The rate of road traffic accidents increased from 32.42 per 10 000 vehicles to 13.20 per 10 000 vehicles.In the past ten years,a total of 7 795 traffic accidents occurred,of them 2 659 minor road traffic accidents,2 817 general road traffic accidents,2 264 serious road traffic accidents,and 55 special serious road traffic accidents.The main reasons for road traffic accidents are wrong use of lanes,retrograde,improper meeting of vehicles,insufficient distance between vehicles and driving license issues.According to statistics,road traffic accidents occurred in 5 200 people in passenger cars and 3 095 traffic accidents in trucks.The most frequent traffic accidents occurred in 2014.Overloading makes traffic accidents more likely to cause fatal traffic accidents.The probability of fatal traffic accidents caused by speeding is about 2.25 times that of non-speeding accidents.We tried to use the 75% of samples extracted from the data as training samples to build a random forest model to fit the nonlinear relationship between meteorological variables and the severity of the traffic accident,and the remaining 25% as the test samples were used for testing.In the random forest model,the optimal mtry parameter value is 5 and the number of decision trees is 400.Wind direction,atmospheric pressure,and temperature may be weighted more heavily than other variables.The error rate of the random forest model is estimated to be 51.09%,while the error rate for the prediction of general traffic accidents is the lowest(45.97%).Similarly,in the neural network model,the calculated error rate is 61.01%.The error rate for minor traffic accidents is the lowest(35.84%).Conclusions This study uses epidemiological methods to describe the distribution of traffic accidents in Suzhou City,Anhui Province,from 2008 to 2017.The real-time meteorological factors were used to predict traffic accidents based on the R software,and to provide reliable information for researchers in the public health and public safety fields.The results of this study show that for the prediction of the severity of traffic accidents using real-time meteorological data,both the random forest model and the neural network model have certain applicability.Future research should include more comprehensive variables and different city data to compare,and the possibility of other models should also be considered. |
PDF Full Text Request