We built a predictive model for a train signalling company that provides warnings of train delays to controllers up to 60 minutes in advance, 50% more accurately than any previous forecasting method.

The model has the potential to save up to 200 minutes of lateness every day at the London terminus.

Customer

A signalling company providing traffic management software to Network Rail controllers as part of the Digital Railway initiative.


Problem

The company currently manages railway traffic on a major route out of London. There are nine control points and 120 miles of rail network on that line operating close to 2,500 trains each day that start or end at the main London terminal. Fewer than 30% of trains are within one minute of their plan at any location. The number of trains arriving at their destination within their targeted scheduled time is currently 82%. The aim is to achieve 90–92%.

In June 2018 the company launched a real-time monitoring system on the route. Faculty was asked to build a layer of intelligence into its product to predict when delays were likely to happen and then to guide controllers to take the right action to minimise the impact on the network.


Solution

We took a three-stage approach to modelling, moving from predicting delays to understanding the underlying causes and finally to recommending actions to controllers.

To predict the emergence of delays at major stations across the network, we used historical data to create snapshots of the state of the network. This data comprised features such as the position of trains, their lateness and the timetable. To capture the complex relationships in the network, we trained two different models, one to learn the linear characteristics of delay (e.g. how late are trains now), and another to learn how interaction across the network builds up to cause delay.

To help controllers understand the causes of delay, we trained models to learn about fundamental patterns in the way the network operates. This included analysing the impact of short-term planned trains to throughput limits at particular junctions.

Finally, we tested the extent to which the historical data could support making recommendations to controllers. We focused on ‘re-platforming’ (sending trains to alternative platforms to those originally designated) as a tangible and common way for controllers to intervene and smooth out problems with delays. At the London terminal, one in five trains is re-platformed. To help optimise this decision we built a model to predict the risk of double occupation, and therefore delay. By training models to predict both the arrival time and the time a train takes to turn around, we were able to create an algorithm that could recommend the optimum platform change.

Each of the models was built into a live dashboard to help controllers manage delays pre-emptively.


Impact

The predictive model provides warnings of delays to controllers up to 60 minutes in advance, 50% more accurately than any previous forecasting method. The re-platforming model has the potential to save up to 200 minutes of lateness every day at the London terminus.