Prediction of Seasonal Variation in Traffic Collisions on Rural Highways Using Neural Network Regression Models: A Case Study in the Province of British Columbia

Abstract

Traffic collisions are one of the world's major problems. According to the World Health Organization (WHO), Over 3,700 people die on the world's roads every day and tens of millions of people are injured or disabled every year. Various tools/methods were developed to assess highway safety including collision rates, linear regression and generalized linear regression methods. Collision prediction modeling is the recommended technique for estimating road safety in the American Association of State Highway and Transportation Officials (AASHTO) Highway Safety Manual (HSM), however the prediction models put a significant weight on the road geometrics (lane width, shoulder width and type, horizontal alignment, and grade, etc.) and Annual Average Daily Traffic (AADT). Traffic seasonal variations and weather effects have not been incorporated in the prediction models. Research has also shown the non-linear relationship between collision frequency and exposure. Previous studies indicate that weather especially winter weather is strongly associated with the traffic collisions and about 24% of all collisions are weather-related. Collison risk usually increases from 50 to 100 percent during precipitation. This study analyzes the seasonal variations of traffic volumes and collisions on rural highways in British Columbia, Canada. Collision risks related to weather factor including temperature and precipitation are investigated and assessed. Neural network regression models are developed to predict the seasonal variation of traffic collisions. This study concludes that the proposed models could incorporate traffic seasonal variations, temperature and precipitations, address no-linear issue and provide over 90% accuracy of traffic collisions over 5 year-period and about 30% more accurate estimate of traffic collisions comparing the prediction model developed by British Columbia Ministry of Transportation and Infrastructure (BCMOTI), Canada. This paper also suggests that the collisions can be more accurately predicted if seasonal variation of traffic volume and weather-related factors can be incorporated in the prediction model. This will be important for Canada and other countries with strong seasonal weather impacts (e.g. severe winter conditions) as the models developed will be able to provide guidance for winter road maintenance and traffic management policies in order to reduce traffic collisions.