AI in public transport

AI in public transport means training a model on data an operator already has: tap-ins and tap-outs, journey data, passenger counts and timetables. You can use it to forecast how many passengers will use a route, see where a service becomes crowded, or plan maintenance and staffing more effectively. It is one of the AI applications we build.

For GVB, we first helped select the data use cases with the greatest impact. You can find more of our work in the sector under mobility and logistics.

With a model that combines historical tap-ins and tap-outs by route, stop and hour with calendar information: school holidays, events, weather and planned works. This lets you forecast demand for a new timetable or a busy period.

The pitfall is that the past is not always a reliable guide. A new route, a diversion or a change in working-from-home patterns affects demand, and the model needs to learn from that. Read more about forecasting.

More than you might think. Operators submit their timetables in NeTEx, current vehicle positions and punctuality data in KV6, and changes such as cancelled services in KV17 to the NDOV data portals. This source data is open data under a CC-0 licence.

Operators also provide monitoring data to public transport authorities under MIPOV, including boardings, alightings, transfers and occupancy. To see how these sources can be combined with your own tap-in and tap-out data, read about our data and AI platform.

Tap-in data misses passengers who do not tap in, while manual counts are labour-intensive and provide only a snapshot. For GVB, we built software that uses computer vision to count passengers in footage from existing security cameras on vehicles. No one views or shares the footage; only the model analyses it.

This gives you insights over a longer period, such as which routes have more passengers travelling without a valid ticket. Read more about computer vision.

By monitoring and retraining it. A new timetable, a closed stop or a change in travel behaviour affects the inputs, making the results less accurate. This is called drift. For GVB, we therefore set up an MLOps workflow in Databricks with reusable templates, so its travel behaviour segmentation model stays up to date over the years.

Read how this works under MLOps. Your team can manage the model, or we can do it through management and continuous improvement.

Tap-in and tap-out records are personal data if they can be linked to an individual passenger. According to the EDPB, pseudonymised data remains personal data under the GDPR. Transdev and Translink’s OVpay privacy notice states that technical payment card data is processed only in pseudonymised form.

Annex III of the AI Act lists AI used as a safety component in critical digital infrastructure and road traffic as high-risk. Your legal or compliance team must assess whether your application falls into that category; we can help identify the relevant uses. See AI Act and Responsible AI.

Yes, if maintenance data is recorded properly. For GVB’s Rail Equipment department, we developed an anomaly detection model that identifies unusual logged hours. It provides insight into productivity, completeness of time records, overtime and absence for each project, helping teams plan projects more effectively.

If you want to anticipate equipment failures, predictive maintenance is the next step. With ProRail, we used this approach to investigate the detection of wheel defects via fibre-optic cables alongside the tracks.

When you mainly need to know how many passengers travelled, rather than how many will travel. In that case, a Power BI dashboard based on your tap-in and tap-out data and MIPOV figures is more useful than a model. A fixed scheduling rule may also be enough if a route follows the same pattern every year.

If stops and routes cannot be linked across your data sources, start with the data foundation. If a simpler solution fits, we will say so.