AI in Aviation. Optimizing Aircraft Turnaround Through Practical AI 

At any busy airport, the turnaround process – the period between an aircraft’s arrival at a stand and its departure – is a tightly choreographed operation involving strict timelines, constrained space, and many interdependent tasks.

Ground crews must unload baggage, clean cabins, refuel, restock supplies, and board passengers – all within a narrow window. A delay in any one of these steps can set off a chain reaction, disrupting schedules across the airport.

Read more: AI in the Aviation Industry: Top 5 Use Cases

Managing aircraft stands efficiently is a complex undertaking. Although stands are assigned based on scheduled times, late departures often lead to prolonged occupancy, causing arriving flights to queue even when other stands appear available.

The real challenge lies in accurately predicting when a stand will be vacated – taking into account the constant flux of real-time operations. In environments where hundreds of aircraft move daily, even small prediction errors can compound into significant operational bottlenecks.

Challenge: Turnaround in aviation

At first glance, optimizing turnaround times may seem like a purely operational goal. However, the benefits stretch far beyond better logistics. Accurate predictions can dramatically improve stand utilization, enabling more aircraft to pass through the airport with fewer disruptions.

This has a direct impact on flight punctuality, airline profitability, and passenger satisfaction.

Moreover, providing a real-time operational view of the airport empowers staff to react quickly to changes. Whether it’s rerouting an incoming aircraft or reallocating ground crews, having up-to-the-minute insight into aircraft movement and stand occupancy turns reactive chaos into proactive control.

Approach: From Vision to Pragmatic Success

The initial vision was ambitious: develop a comprehensive system combining computer vision with data analytics. Camera feeds would automatically detect key turnaround events – fueling, catering, passenger boarding – and feed these visual inputs into a broader data pipeline for near-real-time stand clearance predictions.

The team developed a machine learning pipeline generating stand clearance predictions in real time, with models ingesting streaming data and refreshing predictions every minute. Built on Databricks, this architecture enabled real-time processing, airport system integration, and results sharing through an intuitive dashboard where operators could visualize, assess, and use the predictions.

The Most Remarkable Decision: Choosing “Old School” Over Cutting-Edge

To this point, it all seems like an everyday ML project, not even the most sophisticated, and… from the technical point of view it truly isn’t – and that’s exactly what makes it interesting.

We deliberately chose simplicity over complexity in selecting algorithms, as it turned out that classical, we can say even old-school statistical algorithms, when applied well, deliver matching results at a fraction of the cost compared to the state-of-the-art ones, explains Jakub Berezowski, Data Scientist at Addepto.

The most remarkable decision was NOT pursuing advanced deep learning or transformer models, but reaching out to more modest ones, “boosted” with manually incorporated domain expertise. While advanced models could achieve similar results, they would be significantly more expensive in computational resources, development time, maintenance complexity, and infrastructure costs.

Instead of relying solely on raw event data, we worked closely with operations staff to understand what truly drives turnaround times.

This way, it was possible to develop features, rooted in operator intuition and experience, that gave our models the context they needed to make better predictions. The impact was significant: models using these enriched features doubled their predictive accuracy compared to models using raw data alone.

This approach of combining classical algorithms with expert domain knowledge proved far more cost-effective and pragmatic.

Business benefits: Better, cheaper, faster

One of the most important takeaways from this project was the clear trade-off between technical advancement and operational value. While more advanced models might have squeezed out slightly more accuracy, the cost in infrastructure, maintenance, and complexity would have far outweighed the benefit.

For example, a deep learning model might predict a 45-minute turnaround with high confidence that the error wouldn’t exceed 4.5 minutes, while our simpler model provides reassurance of a 5-minute margin. That half-minute gain would come with a 10–100x increase in resource consumption. From a business perspective, the simpler solution wasn’t just good enough – it was exactly right.

Additionally, deploying on Databricks streamlined the entire process, allowing us to reuse existing data pipelines, connect with other teams, and rapidly iterate without reinventing infrastructure from scratch.

Key Insights

While event data from external vision systems didn’t prove as impactful as initially hoped, the project delivered tremendous value. The most predictive variable turned out to be airport congestion, not individual turnaround events. This insight shifted focus from micro-events toward broader operational context.

Conclusion: Practical AI that works

This case study highlights a powerful lesson: the best solutions are not always the most complex. By leveraging expert knowledge, choosing interpretable models, and focusing on real-world constraints, we built a system that solved a real business problem without over-engineering it.

Explore our Case Studies
to discover how our team delivers actionable, cost-effective AI and analytics solutions across industries.