Predicting the function recovery time for railway incidents

Abstract

When an incident is reported at ProRail, many decisions have to be made. All those decisions influence the function recovery time (FHT), which is the time span between the moment an incident is reported and the moment everything is restored. When the FHT can be accurately predicted, better decisions will be made with regard to incident handling, and delayed train passengers can be accurately informed. The aim of this research is to build a supervised machine learning model to predict the FHT. Currently, a Decision Tree is implemented at ProRail to produce an initial prognosis. Ideally, our model will be suitable for both generating an initial prognosis, and updating this prognosis during the recovery process. Historical incident data of ProRail will be used to train the model on. This data is extended by adding a few other data sets, to increase the number of relevant variables. The resulting data set is divided by incident type, because the variables that correlate with the FHT are very different between incident types. For each type, a different set of features is selected, and a different model is trained. The models have to meet three requirements. First, all of our variables are or can be discrete, so the model should be able to work with categorical data. Second, when the initial prognosis is made, usually the values of only a few variables are known, so the model should be able to handle this, and also be able to make a better prediction based on more values later in the process. Third, the model will be used to support humans in making choices, so the output of the model should be easily interpretable for a human. As a result, Bayesian Networks and the k-Nearest-Neighbors algorithm are chosen to be implemented. These algorithms are tested on data sets for rolling stock incidents, section TOBS incidents and collision/hindrance incidents. The mean distance between the predicted FHTs and the actual FHTs is similar for both algorithms; between 45 minutes and one hour for all three data sets. This performance is compared with the predictions of the currently implemented Decision Tree, using a small test set. On this set, our models do perform similarly or better.