A neural network (NN) model to predict intersection crashes based upon driver, vehicle and roadway surface characteristics

In this paper, a neural network (NN) model was developed to predict intersection crashes in Macomb County of the State of Michigan (MI), USA. The predictive capability of the NN model was determined by grouping the crashes into these types: Fatal, injury and property damage only (PDO) () accidents. The NN approach was used to develop and to test multi-layered feedforward NNs trained with the back-propagation algorithm in order to model the non-linear relationship between the crash types and crash properties such as time, weather, light and surface conditions, driver and vehicle characteristics, and so on. 16000 cases of the crash data were used to train the NN model and the model testing was done by another set of 3200 crashes. A sensitivity analysis was performed to define the effect of crash properties on the crash types. The approach adapted in this study was shown to be capable of providing a very accurate prediction (90.9%) of the crash types by using 48 design parameters (selected based on statistical significance among crash properties defined in the data file). The results were considered to be very promising and encouraging for further research by the expanded data sets in order to estimate future year dependent variables with the model built.