Detecting probability of ice formation on overhead lines of the Dutch railway network

Probability of ice formation over the railway network

Ice formation along the overhead lines is a potential source of infrastructure failure and transportation disruption. Modelling ice formation over structures is challenging because it may occur at a very local scale, often far away from measuring stations. Experts in the field have empirically defined models of ice accretion under laboratory conditions. However, the significant number and complexity of the geophysical parameters included in these models, poses major hurdles for their application at the national level. In this project, we propose a data-driven solution overcoming such hurdles, only stemming from the ice formation observations that the Dutch railway infrastructure collected during six years.

Data analysis

We applied a process of feature engineering to the observations, in which we enriched the ice formation observations with weather variables (temperature, relative humidity, dew point). Then, we trained a Gaussian Process (GP) with observations in a binary classification set up, so the model could learn to distinguish between a “normal” day and a “frost” day. After calibrating the GP, we applied the model to each day during the study period at the national level, so we obtained daily maps showing the probability of developing frost over rail structures. The overall accuracy of the model shows that we can predict 70% of the “frost” days.

Additional information

  • Funding: KNMI/Pro-Rail
  • Duration: Feb 2018 - Oct 2018
  • Contributions:
    • IEEE eScience 2018 conference
    • de Vos, M. G., Hazeleger, W., Bari, D., Behrens, J., Bendoukha, S., Garcia-Marti, I., van Haren, R., Haupt, S. E., Hut, R., Jansson, F., Mueller, A., Neilley, P., van den Oord, G., Pelupessy, I., Ruti, P., Schultz, M. G., and Walton, J.: Weather and Climate Science in the Digital Era, Geosci. Commun. Discuss.,, in review, 2019.
Irene Garcia-Marti
PhD Data Scientist

I have a keen interest in applying machine learning methods in the field of spatio-temporal analytics.