Researchers trained an AI to sense when a flying wind kite is about to snap — and it works with striking accuracy
Hundreds of meters above the ground, a large kite slices through the wind, its tether pulling taut to drive a generator below. It looks controlled — but inside that tether, invisible force spikes can surge without warning, overwhelming actuators and quietly wearing down the kite’s structure with every undetected event.
These sudden load peaks have long been one of the hardest problems in airborne wind energy: difficult to anticipate, costly to ignore, and until recently, impossible to flag before the damage is already underway. A one-second warning may not sound like much — but for an autonomous system flying at altitude, it could be enough to act.
What are critical load events — and why do they matter?
In airborne wind energy, a critical load event is a moment when tether tension shoots past a threshold set by the winch controller. That threshold marks the boundary where the system starts to lose control.
The research team used Kitepower’s leading-edge inflatable V9 kite as their test platform — a real, operational system rather than a simulation.
When tension exceeds that limit, actuators can saturate — hitting their physical limits and losing the ability to steer the kite effectively. Structural components then absorb stress they weren’t designed to handle repeatedly. Each undetected spike adds wear, and over hundreds of flight cycles, that cumulative degradation threatens commercial viability. Without a human pilot in the loop, real-time prediction isn’t just useful — it’s essential.
A dataset built from more than 1,000 flights
The research team used Kitepower’s leading-edge inflatable V9 kite as their test platform — a real, operational system rather than a simulation. Their dataset came from the system’s SCADA logs: more than 1,000 pumping cycles, covering 26 hours of kite operation across a wide range of wind speeds and conditions.
From that data, researchers detected tether force peaks and extracted “pre-peak” time windows — slices of sensor data capturing system behavior in the seconds just before each threshold exceedance. Features were computed across statistical, temporal, and spectral categories, drawing from sensors on the kite, tether, and ground station alike.
How the machine learning model learns to see danger coming
The core approach is supervised classification. Shown examples of both normal operation and pre-peak dynamics, the model was trained to tell them apart — picking up on subtle patterns in sensor signals that precede a force spike.
What it “sees” is a stream of multivariate sensor readings in the seconds before a potential event, tracking how multiple measurements evolve together and flagging combinations that signal trouble ahead. The prediction horizon sits at one second — enough time for an autonomous flight controller to issue a corrective command. The model was tested on flight data it had never encountered during training, which is what makes its performance operationally credible.
85.4% recall and a 90.9% F1-score: what the numbers mean
The headline results are an 85.4% recall rate and a 90.9% F1-score. Recall tells you how many actual critical events the model caught. At 85.4%, it correctly flagged the vast majority of dangerous load spikes, missing roughly one in seven.
The F1-score balances recall against precision, penalizing both missed events and false alarms. At 90.9%, the model proves selective as well as sensitive — catching most of what matters while staying quiet the rest of the time. That’s exactly the behavior you want from an early-warning tool. The extracted features also revealed patterns in system dynamics that gave researchers new insight into how kites behave under stress.
From prediction to prevention: what comes next for airborne wind energy
A one-second warning is only valuable if something acts on it. The logical next step is weaving this model into flight controllers directly, so a predicted load spike automatically triggers a preventive maneuver — adjusting the kite’s trajectory or easing tether tension before the spike arrives.
Fewer uncontrolled load events means less structural wear and longer-lasting hardware. For an industry still working to prove its commercial case, those gains are foundational. Real-time integration will demand low-latency inference and robust failure modes, and whether the model generalizes across different kite designs or wind regimes remains an open question. Still, with a working prediction framework already showing strong accuracy on real flight data, the path from early warning to active prevention is considerably shorter than it was before.
Learn more about this discovery here: Pasquali, M., Cayon, O., Theodoulis, S., Farges, C., Airimitoaie, T.-B., Lanusse, P., and Schmehl, R.: Machine learning-based prediction of critical load events in airborne wind energy systems, Wind Energ. Sci. Discuss. [preprint], https://doi.org/10.5194/wes-2026-180, in review, 2026.
Daniel García is an Editor-in-Chief with strong expertise in structural work and engineering principles. He combines this technical foundation with deep knowledge of energy, spatial design, and emerging technologies, bringing a forward-thinking and analytical approach to editorial leadership.
