Identifying risky locations for road accidents due to crosswind
Crosswind is a potential source of road accidents, especially for unloaded trucks. We joined forces with colleagues from the ILT and RWS datalabs to assess how feasible it is to build predictive models capable of identifying risky locations for truck accidents. In this way, it is possible to proactively plan traffic inspections that might reduce the probability of accidents.
We combined a given set of locations of truck accidents with road characteristics, traffic intensity, and weather variables. The location of the accidents cofnorms a positive class only, so we generated a negative class considering the spatio-temporal structure and intensity of traffic, making sure the classes are difficult to separate. After this step, we trained machine learning models from the ensemble learning family (i.e. Random Forest, Gradient Boosting) in a binary classification set up. We obtained a moderate overall accuracy, suggesting the models are capable of predicting, to some extent, the probability of truck accidents.