In a large passenger airport, a delayed aircraft can cause a number of repercussions for other flights and ground staff must react quickly to avoid major delays. We were asked by a large commercial aerospace company to investigate whether the task of assigning gates to aircraft could be automated using machine learning. Our clients provide software that facilitates manual management of gate allocation which can be semi-automated using rules.

We were asked to build a model that can predict which gate to allocate to an aircraft based on historical logs of flights in and out of an airport.


We initially spent some time investigating features that characterise particular aircraft and gates such as wingspan, number of passengers etc.  We tried a number of simple classification models using algorithms like Random Decision Forests and had moderate success. Eventually we used a recurrent neural network approach to treat the list of flights as a sequence of events and maintaining airport state between flights.

We also took into account that flight volume and delay frequency are seasonal variables. Delays often snowball in cold conditions when flights are held up due to inclement weather.

We were able to train a model that correctly guesses which gate to park a flight at 80% of the time*

We produced a demo web application which the client was able to re-train by uploading a CSV file containing historic flight information and could use to make new predictions via a REST API.

The project was carried out over 4 weeks.

* the model correctly predicted the gate within the top 3 results