Case study: Reducing train delays using highly accurate forecasting
Every day, 127,000 commuters travelling into London have their train delayed or cancelled. To reduce these delays, controllers have to quickly move trains between platforms to keep traffic moving. However, slow and poor decision making often increases delays further, having knock-on effects for transport links across the country.
In 2018, Faculty built a model that could predict delays at one of London’s most popular rail stations and gave recommendations on how to reduce them. Faculty’s forecasting technology has been shown to save over 3 hours worth of delays at the London terminus every day.
Faculty worked with a signalling company that provides traffic management software to Network Rail controllers. They manage railway traffic on a major route out of London.
To reduce train delays, controllers use software that tracks trains arriving and leaving the terminus to determine which trains should go on which platforms. But these legacy systems only tell them what’s happening now, leading to rushed, ineffective decision making that causes knock-on delays.
Most train delays are near-impossible to prevent; damaged rails, trespassing and slippery leaves on tracks can’t be forecasted. Train controllers have to respond to incoming reports of delayed and cancelled services by re-platforming trains to keep the traffic flowing and smooth out the service. These time critical decisions are complicated, requiring controllers to be aware of all of the trains arriving and leaving.
Old traffic management systems are also burdened by slow data feeds and a lack of data integration, only telling controllers what was happening in real-time. Staff were, therefore, unable to prepare effective interventions before the traffic hit.
An ineffective or slow decision can have knock-on delays that last up to 24 hours, affecting transport links nationwide and costing train operators tens of millions in compensation. The signalling company in question wanted to establish whether train delay forecasting could increase the number of trains arriving at their destination on time from 82% to 90-92%.
Faculty built a ‘layer of intelligence’ into the signalling company’s real-time monitoring system. The solution was able to forecast delays an hour in advance and helped controllers decide on the best interventions to reduce delays.
To predict delays at the London terminus and other major stations across the network, Faculty used historical data including train position, lateness and scheduled arrivals to build a picture of the entire network. Faculty trained two different models: one to learn the characteristics of delay (e.g. how late are trains now), and another to learn the causes of delay to generate highly-accurate forecasts.
To help controllers understand the causes of these delays, Faculty trained models to learn about fundamental patterns in the way the network operates, such as the impact of trains added at short notice, commercial fleets and throughput limits at particular junctions.
Faculty then created an algorithm to optimise re-platforming. The model predicted both the arrival time and the time a train takes to ‘turn around’, before recommending the optimal platform change for controllers to action. Controllers no longer needed to waste time calculating the best position for a train, or risk making a wrong decision.
Each model was built into a live, intuitive dashboard to help controllers quickly view future delays, understand what was causing them, and decide how best to intervene.
Faculty’s forecasts accurately generated train delay predictions an hour in advance, allowing controllers to make earlier decisions when needing to re-platform a train. This technology could be deployed to stations across the country to reduce train delays nationwide.
The predictive model provides delay warnings to controllers with far greater accuracy – by more than 50% compared to any previous train delay forecasting method. Faculty’s layer of intelligence guarantees every re-platforming decision is optimal, executed rapidly, and made in advance, minimising delays before they affect other services. The re-platforming model can save up to 200 minutes of lateness every day at the London terminus.
more accurate than previous forecasts
mins of delays saved
min pre-warning of delays
To find out more about Faculty's work, visit their website.