Application of Simulation-assisted Machine Learning for Yard Departure Prediction

Increasing the modal share of rail freight is an ongoing goal in Europe and North America. Yards can play an important role in realizing this target by their reliable and predictable performance. We aim at predicting yard departures by implementing a simulation-assisted machine learning model via two general and step-wise concepts for including the predictors. The former adds all predictors at once, and the latter adds them per the availability or the sub-yard. The data used for training the model is a one-year real-world operational data set from a European hump yard and multiple two-year simulation data sets from a representative hump yard in North America. To the best of our knowledge, no previous research has attempted to implement a generalizable prediction model between the European and the North American contexts. The model is developed on a decision tree algorithm based on a 10-fold cross-validation process. Comparing the model performance on three data sets: the real-world, a baseline simulation, and an ultimate randomness simulation shows that the model has a similar performance in the first two data sets with a respective R-squared of 0.90 and 0.87, which shows high capturing of the variance in the data. However, adding large randomness in the simulation decreases the R-squared to 0.70. Results for the step-wise inclusion of the predictors are different for the real-world and simulation data. For the former, adding more operational predictors does not change the model performance, whereas for the latter, adding departure yard predictors increases the R-squared substantially. The global feature importance shows that for the real-world data almost all predictors contribute to a great extent to the predictions, with maximum planned length, departure week day, and the number of arriving trains as the most contributing ones, whereas for the simulation data, the departure yard predictors provide the largest contribution.
QC 20230517
FR8RAIL III
PRATA
Shift2Rail