The research, led by Monika Pasquali of the University of Bordeaux and TU Delft, together with colleagues including Roland Schmehl of TU Delft and control engineers from the Bordeaux laboratory IMS, takes a deliberately data-driven route. Rather than building a physics model of the kite from first principles—an approach that struggles with the chaotic, gusty reality of the atmospheric boundary layer—the team let the data speak. They mined a large-scale SCADA dataset recorded by Kitepower’s leading-edge inflatable V9 soft kite, a commercial-grade system operating with a pumping cycle: the kite reels out under power, generating electricity, then is reeled back in to start again. The dataset spans more than 1,000 pumping cycles and 26 hours of continuous autonomous operation, covering a wide range of wind speeds and operating conditions.

The technical challenge hinges on defining exactly what counts as a dangerous event. The researchers defined critical load events as tether force peaks—instances in which the measured force exceeds a threshold set by the winch controller itself. Because the winch controller is the component that must manage the load, its own limit provides a physically meaningful yardstick rather than an arbitrary cutoff. The team then detected every such peak in the 26-hour record and extracted the time windows immediately preceding each exceedance. These pre-peak windows capture what the kite, tether, and ground-station sensors registered in the moments before trouble, essentially freezing the system’s behavior in the run-up to each event.

From these windows, the researchers computed three families of features: statistical features capturing the distribution of sensor values, temporal features describing how measurements evolve over time, and spectral features revealing oscillatory content in the signals. Importantly, the features were drawn not from a single sensor but from multiple measurements across the entire system—the kite itself, the tether, and the ground station. This multi-source approach matters because a looming force peak rarely announces itself in one channel alone. Subtle shifts in flight dynamics, tension fluctuations, and winch behavior combine into a signature that a single measurement could never reveal.

With these features in hand, the team framed prediction as a supervised classification problem. The model’s job was to look at a snapshot of the system and decide: is this normal operation, or is this the kind of pre-peak dynamic state that ends in a threshold exceedance one second later? Training a classifier on labeled historical flight data allows the algorithm to internalize the patterns that human engineers might only notice in hindsight. The authors report that the extracted features also provide valuable insight into the system dynamics preceding these events, meaning the model is not merely a black box but a lens onto the kite’s behavior.

The performance numbers are striking. On flight data that the model had never seen during training—a crucial test of generalization in a domain where conditions change constantly—the classifier correctly identified 85.4 percent of the critical load events, a measure known as recall. It achieved an F1-score of 90.9 percent, a metric that balances correctly detected events against false alarms and indicates that the model is not simply crying wolf. In an operational setting, that balance is everything: a system that misses real load peaks is unsafe, while one that floods operators with false warnings would erode trust and trigger unnecessary defensive maneuvers that sacrifice energy production.

Why does a single second of warning matter so much? In airborne wind energy, one second is an eternity of control opportunity. A prediction delivered one second ahead of a force peak gives the winch controller time to take preventive action—adjusting reel-out speed, modulating tether tension, or steering the kite away from a trajectory that is building toward overload. Without prediction, the controller can only react after the force has already exceeded its limit, which is precisely when actuators saturate and steering authority is lost. The authors emphasize that the obtained results enable preventive control actions, converting a safety problem from an unavoidable consequence of gusty winds into a manageable control task.

The broader significance lies in what this means for the maturation of airborne wind energy as an industry. Companies such as Kitepower, a TU Delft spin-off, have flown systems for thousands of hours, but insurers, regulators, and grid operators demand quantified reliability before such technology can scale. Every avoided load event extends the lifespan of the kite, the tether, and the winch—the three components where wear accumulates fastest. A data-driven early-warning layer, trained on the system’s own operational records and refined continuously as more flights accumulate, offers a path to autonomous operation that degrades gracefully rather than catastrophically. Notably, the team has released the model code as an open-source repository, awe-event-predictor, and published one of the underlying flight datasets on GitHub, lowering the barrier for other research groups to build on the work.

The study is a collaboration across borders and disciplines, pairing the aerospace expertise of TU Delft’s Faculty of Aerospace Engineering with the control systems and signal processing strength of Univ. Bordeaux, CNRS, Bordeaux INP, and IMS in Talence, France. That combination is telling: airborne wind energy sits at the intersection of aerodynamics, mechatronics, and control theory, and taming its load dynamics requires all three. The preprint is currently under open peer review at Wind Energy Science, with the discussion open to community comments, meaning the reported results will face scrutiny from specialists before final publication.

Still, the one-second horizon invites the next round of questions. Can the warning window be stretched further—two or three seconds—giving controllers even more room to maneuver gracefully? Can similar models transfer between kite designs, or must each system learn from its own flights? And as forecasting becomes preventive control, how should autonomous systems weigh the energy cost of avoiding a load peak against the structural benefit? What Pasquali and colleagues have demonstrated is that the information needed to anticipate danger is already flowing through the sensors of a flying kite; it simply required the right statistical machinery to read it. For a technology whose fate depends on proving it can fly safely through turbulent air for decades, that reading is a milestone worth watching.

Subject of Research: Machine learning prediction of critical tether force peaks in airborne wind energy systems

Article Title: Machine learning-based prediction of critical load events in airborne wind energy systems

Article References: Machine learning-based prediction of critical load events in airborne wind energy systems. (n.d.). https://doi.org/10.5194/wes-2026-180

Image Credits: AI Generated

DOI: 10.5194/wes-2026-180

Keywords: airborne wind energy, machine learning, tether force peaks, Kitepower, wind energy, predictive control, supervised classification, soft kite, structural loads, renewable energy, SCADA data, fault prediction