Predicting road traffic accidents using artificial neural network models

Abstract

As of 2015, Switzerland’s road network was among the safest when compared to other European countries. Nonetheless, the endeavour to further decrease the number of traffic accidents and road deaths remains part of the federal agenda. A proper assessment of the relevant risks is, therefore, of utmost importance. This paper presents a methodology for establishing an accident risk prediction model, which can be used as a decision-making tool in infrastructure management. The methodology allows for an appropriate handling of the available data, examines how it can be used to develop models using artificial neural networks (ANNs) and establishes a systematic ANN optimisation process to determine the optimal architecture of the ANN model. The methodology is implemented using data for accident counts on the Swiss national roads from 2009 to 2012. It has been found that ANNs can be used as a viable method to predict the frequency of road accidents. As accident occurrences are relatively rare events, the data are characterised by a large portion of zero observations. This poses a challenge for the training of the ANN. The results show that such models provide reliable results as indicated by the symmetric mean absolute percentage error, ranging from 17·5 to 32·7%.