There occur 300 – 400 disruptions per day in the SBB railway network. Many of these disruptions lead to train delays. 
SBB uses an application to coordinate the resolution of a disruption to the network, where various stakeholders involved in resolving the disruption communicate with text messages about the state of the disruption and send teams of workers on site to fix the issue. 
We developed and trained different Machine Learning models that can forecast the disruption duration based on the un-structured text messages obtained during the resolution of a disruption.
We leverage LSTMs in combination with word level features to encode the stream of text messages in the model. This allows us to train an end-to-end predictor that can estimate the disruption duration each time a new message arrives. 

Slides

Time: 11:15 - 11:45
Track:

Track 5
organised by SSS


Speaker: Gabriel Krummenacher & Beat Wettstein