A wind turbine controller that simultaneously chases higher revenue and lower wear sounds like an engineering tradeoff with no clean answer. Researchers at the Technical University of Munich may have found one.
In a study published in Wind Energy Science, Anand and Bottasso describe an adaptive economic nonlinear model predictive controller that embeds offline-trained neural network corrections into the turbine’s internal prediction model. Tested in simulation on the NREL 5 MW reference turbine, the approach yielded up to 30% higher economic profit and measurably reduced actuator wear compared to a non-adaptive baseline controller.
What the study found
The headline result is straightforward: the adaptive controller, called ENMPCaug, delivered 9% higher economic profit than the non-adaptive baseline (ENMPC) under standard wind estimation conditions. That gap widened considerably when better wind data was available.
When researchers fed the controller LiDAR-derived wind speed estimates instead of a simple rotor-effective wind speed calculation, the profit advantage over the baseline rose to 30%. Pitch travel and torque travel—measures of how hard the blade pitch actuators and generator drivetrain are working—fell substantially at the same time. Less actuator movement suggests lower wear, though the authors note they did not quantify the resulting maintenance savings due to a lack of reliable component-level data.
Tests ran in closed-loop simulation against an OpenFAST high-fidelity model of the NREL 5 MW reference turbine. Six turbulent wind seeds at 11 m/s provided statistical coverage of fatigue uncertainty across a standard 10-minute evaluation window.
Why a standard reduced-order model is insufficient
Model predictive control works by repeatedly solving a short-horizon optimization problem. To do that fast enough for real-time use, it relies on a simplified internal model—a reduced-order model, or ROM—rather than the full physics of the actual turbine.
The gap is significant. The NREL 5 MW plant modeled in OpenFAST has 33 system states covering blade bending, tower dynamics, drivetrain torsion, and actuator dynamics. The ROM used in this study has only 8 states. That simplification is necessary for speed, but it means the controller is making decisions based on predictions that do not fully capture what the turbine is actually doing.
This discrepancy is called a plant-model mismatch. When the controller’s internal picture of the turbine diverges from reality, its choices of pitch angle and generator torque become suboptimal—the turbine may sacrifice revenue or accumulate more fatigue than necessary. Existing fixes such as gain tuning or state observers can partially compensate, but they cannot fill in physics the ROM simply does not represent. Gain tuning, in particular, may produce non-physical results when the underlying model structure is wrong.
How the adaptive controller works
The core idea is to teach the ROM to correct its own errors before the controller ever runs in the field. A feed-forward neural network with one hidden layer of 20 neurons is trained offline on synthetic data generated by high-fidelity OpenFAST simulations, learning to predict the difference between what the ROM would forecast and what the plant actually does.
That correction term is added directly to the ROM, producing an augmented model called ROMaug. It tracks actual turbine dynamics more closely across a wide range of wind speeds and operating conditions. In open-loop testing, the augmented model reduced both the mean and standard deviation of prediction errors for all eight system states.
Incorporating fatigue costs into the optimization required a separate technical solution. Cycle counting—the standard engineering method for estimating fatigue damage—is inherently discontinuous, which breaks gradient-based solvers. The researchers address this using parametric online rain flow counting (PORFC), converting the discontinuous counting process into a continuous, differentiable expression compatible with the numerical optimization at the controller’s core. Because all adaptation happens offline, the corrected model can be fully verified before deployment—a meaningful practical advantage over online learning, where safety guarantees are harder to establish.
Computational cost and real-time feasibility
The performance gains come with a computational price. Running ENMPCaug with 10 sequential quadratic program (SQP) iterations requires more than double the CPU time of the non-adaptive baseline at the same iteration count, with the extra cost coming from evaluating the neural network correction at each step.
The trade-off becomes more favorable when iteration counts are reduced. ENMPCaug running at only five SQP iterations still achieves 22% higher economic profit than the baseline at 10 iterations, while adding just 15% more computational effort. Fewer iterations are needed because the better internal model converges to a good solution faster. The authors acknowledge the current implementation was not optimized for speed, suggesting that high-performance computing hardware combined with targeted software optimization could make the controller real-time feasible—though this remains to be shown.
Scope, limitations, and next steps
The study’s economic objective covers only one component: tower-base fatigue. Bearings, the drivetrain, and pitch bearings are explicitly identified as components requiring dedicated fatigue models in future versions of the controller.
The profit formulation also does not connect fatigue loads to failure probabilities or operation-and-maintenance costs. The authors attribute this to a lack of validated data linking loads to failure rates, rather than a modeling choice, with component costs amortized as a proxy instead. Extensions outlined for future work include online parameter tuning alongside the existing offline augmentation, variable electricity prices evaluated over longer time horizons, and testing across a broader range of inflow conditions and installation sites. Whether the adapted model generalizes across different sites — where turbine behavior may differ — has not yet been investigated.
Taken together, the study demonstrates a measurable and consistent profit improvement from neural-network-corrected model predictive control, with the largest gains appearing when higher-quality wind information is available. The main technical contributions are the PORFC fatigue formulation, the offline augmentation approach, and the finding that a simpler five-iteration configuration preserves most of the economic benefit at a modest computational premium.
Kelly is an experienced writer with 15 years of experience exploring the big stories that shape our world, from tech breakthroughs and space exploration to climate, energy, and the fascinating quirks of science. She has a talent for turning complex ideas into sharp, memorable insights that stay with readers long after they’ve finished reading.
