To reduce the impact of delays on the operation of KLM Royal Dutch Airlines, a Xomnia X-Force team designed and implemented one predictive model and revised and improved another, going from problem identification to production environment in under three months.
Following this collaboration, KLM is now capable of proactively acting on delays. As a result, delays have less impact on the flight schedule, which leads to lower delay minutes throughout the day and less cancelled flights. Ultimately, the outcome is more efficient operations and higher customer satisfaction.
The main challenge that we see with data science at companies nowadays is identifying the right problem that actually adds value to the business. The second challenge is developing a mature way of working with the right skill set to bring a concept all the way to production.
Robin van den Brink, X-force analytics translator
Case: Using data to predict delays
Based in Schiphol Airport in the Netherlands, KLM Royal Dutch Airlines operates flights to over 160 destinations worldwide. As the company celebrates its 100 year anniversary, it aims to remain a front runner in a competitive industry by becoming more data-driven.
One of the challenges that KLM seeks to mitigate through data is delays, which have a huge day to day impact on operational processes and customer satisfaction. Several processes have to come together at the same place and time for a smooth operation, for example, ground processes (like cleaning and fueling of the plane), baggage services, and the passenger flow. An unexpected hiccup can cause propagated delays that can be seen throughout the schedule of the entire day. By predicting the delays the uncertainty of delay can be used to optimize the schedule.
To achieve this, an X-Force team from Xomnia made up of an analytics translator, two data scientists and two data engineers was assigned to work on researching the repercussions of delays at KLM Royal Dutch Airlines. In order to determine which solution would have the most effective action for day-to-day operations at KLM, the team started by analyzing the problem and found two opportunities to make the most impact
Solution: Updated and new delay prediction models
The first area of development was found to be improving a preexisting delay prediction model that KLM had been using to predict delays one day before a flight. The second solution was focused on predicting delays throughout the day of operations. Both solutions added intelligence to the operational tooling that KLM developed, which enabled to optimize schedules in a smarter way.
Using information about every single flight, involving the conditions at Schiphol Airport and foreign airports - such as the weather forecast and historic delays - the X-force team trained a deep neural network to predict the delay of a flight. During the development process, a dashboard was written in Python and Dash, and developed to a Kubernetes cluster. The dashboard was developed to help end users understand the predictions and verify the outcome based on years of domain expertise.
For optimal results, the data-driven tools not only need predictions but also confidence in all predictions. This means that the deep learning architecture is customized to generate a distribution of the expected delay. Using a probability distribution instead of a single prediction value, the operational tools of KLM can minimize the expected cost while at the same time reducing the impact on scheduled processes by avoiding unnecessary rematching of airplanes to planned flights. In practice this means that the tool takes the delays into account and checks if another airplane should take over or that the delay has a low impact.
Just like the dashboard, the model is also deployed to a Kubernetes cluster. Using the Atlassian CI/CD stack, improvements to any component of the model can be tested and deployed in a matter of minutes.
Within 3 months, the X-Force delivered a revised, retrained and significantly improved the pre-existing model and deployed it in the new Kubernetes architecture. In addition to this, a new model was created in three sprints that is capable of continuously predicting delay for all the (upcoming) flights during that day.
At the time of writing, two prediction models produce a combined 18,000 predictions per month to support the day to day operations of KLM and feed their operational optimization tools. KLM is now better able to mitigate the consequences of delay and act proactively which prevents the forced cancellations of flights. They can proactively swap planes, reschedule flights, rebook passengers, and prevent people from having to stay overnight in Schiphol.
Indirectly, this means happier customers, and a future-proof system that can cope with delays without it having a huge impact on the day-to-day operations.
KLM has asked the X-Force team to continue working on the two predictive models and expand the scope on all of the processes around aircraft to identify opportunities to make the operation more predictable. The X-Force way of working has become the new standard within the entire Prediction team of KLM. This was realized with a close intertwined collaboration between the X-Force and the Prediction team, with several workshops, presentations and hands-on support.