With up to half a million journeys each day, delays are inevitable in train operations. As they occur, information is crucial to travellers planning their day. The Danish State Railways, DSB, sought a more accurate way to estimate and predict delays.

Large amounts of data and various details from delay to delay made it difficult to predict exactly how much time a specific issue – a damaged catenary system, a malfunctioning train door, or a maintenance problem on the tracks – would cost travellers on their journey.

The existing system was a relatively simple projection model that estimated how much time had already been lost during a delay. The delayed train would then try to catch up with the delay, resulting in an optimistic model that had potential to integrate AI in a new way.

Emil Møller Marker, VP for strategic AI at DSB, describes that the company wanted to change their optimistic expectation of catching up delays to a realistic approach powered by an AI model.

»We wanted to collect, analyse, and act on data to predict delays more efficiently. DSB prioritises customer satisfaction and we serve a lot of customers each day. That is why we wanted to provide the most accurate information possible.«

We joined forces with DSB to create a tool that analyses not only singular events but also the bigger picture. Using real-time data, historical patterns, and contextual information, a swarm of AI models distributed nationwide was trained to estimate delays more accurately.

Instead of using a single large model, the swarm allowed DSB to implement the system one railway line at a time, addressing one issue without disrupting the models already in use.

The AI swarm is orchestrated by our PULSE platform. The platform monitors every train in real time, cleaning and processing train data before feeding it to the AI models making delay predictions. Simultaneously, PULSE tracks the AI swarm’s performance to ensure predictions are as accurate as possible.

While the system predicts train delays in a much more efficient way, ongoing enhancements will ensure even more timely information reaches every customer.

Drift, for instance, is a common feature of AI models. New information might emerge, or the circumstances in which the models work could change. However, the design of the AI swarm ensures that the necessary training of one model does not interrupt the rest of the system.

Mathias Engel, AI and data manager at Netcompany, advised DSB to adopt an experimental approach. This meant starting at a short but significant line to prove they had the right solution.

The first test was placed on Kystbanen, one of the most heavily used regional lines in the country with a lot of commuters using it each day.

»We needed to demonstrate that our solution worked on a line with a lot at stake. That is why we chose Kystbanen. Once we proved our solution was effective, we could implement it to another line. Another success meant yet another line was added,« Mathias Engel says.

As a train travels across the country, it passes through several models within the AI swarm along its route. Between each model, real-time forecast data is exchanged. Picture a handshake between each model initiated by the passing train. Mathias Engel puts it shortly:

»We can cover a train journey from start to finish with several models. Our system is more realistic, because it is powered by AI, and it takes how the train has already run into account before calculating if it can catch up with a delay.«

The overall goal to improve predictions was quickly achieved. From that point, increasing precision became the main objective.

»When you leave your home and check your navigations app, the DSB-app, Rejseplanen, or the monitors at the train station, you will know when your train arrives within three minutes with 95 percent accuracy«, says Mathias Engel.

The closer the train gets to a station, the easier it is to predict a delay and whether the train can catch up on it or not.

But, says Mathias Engel, the predictions are made in real time in all of Denmark with a high level of quality.

»We have built a flexible system that can be rolled out across the broader logistics and transport sectors. It does not have to be trains; it could be buses or a delivery service. In any business with a plan that risks going awry, our models and system can be used.«