A new controller for wind turbines can raise economic profit by up to 30 percent — not by squeezing out more power, but by teaching the turbine’s internal model to recognize its own blind spots.
Researchers Anand and Bottasso at a German university, funded by the BMWK e-TWINS project, have published a study in Wind Energy Science presenting an adaptive economic non-linear model predictive controller for wind turbines. Tested in simulation against the NREL 5 MW reference turbine, the system uses a neural-network-corrected internal model to simultaneously maximize revenue and minimize fatigue damage costs — while measurably cutting pitch and torque travel compared to a non-adaptive baseline.
What Was Developed and Published
The study introduces an adaptive economic non-linear model predictive controller (ENMPC) for wind turbines. It pairs a reduced-order physical model with a neural-network correction term trained offline on high-fidelity simulation data. Together, they form an augmented internal model — called ROMaug — that tracks real turbine behavior more accurately during operation. Funding came from the German Federal Ministry for Economic Affairs and Climate Action through the e-TWINS project.
The core objective is economic: maximize profit by balancing electricity revenue against the cost of fatigue damage, particularly at the tower base.
Why a New Approach Was Needed
Standard MPC controllers use reduced-order models to keep computation fast enough for real-time use. These models approximate turbine dynamics with only a handful of degrees of freedom, which inevitably creates a gap between predicted and actual behavior — a plant-model mismatch. That gap has real consequences, pushing control decisions away from the economic optimum and nudging the turbine toward operating outside desired limits.
Previous adaptive controllers addressed this only partially, tuning gains through simplified linear models without explicitly correcting modeling errors. Earlier fatigue-aware MPC approaches either relied on indirect proxies for fatigue or lacked any mechanism to compensate for inaccurate predictions. This work targets both problems at once.
How the Controller Works
At the heart of the system is a feed-forward neural network with a single hidden layer of 20 neurons. Trained offline, it predicts the error between the reduced-order model and a high-fidelity OpenFAST plant model across a wide range of operating conditions. That trained network is then added as a correction term to the ROM, producing the augmented model used inside the controller.
Fatigue costs are computed directly within the optimization using the parametric online rain flow counting (PORFC) method, which converts the inherently discontinuous process of cycle counting into a continuous, differentiable formulation — making it compatible with gradient-based optimization. State and wind speed estimators supply accurate initial conditions and disturbance inputs at each 100 ms control step.
Key Results from Simulation Tests
Open-loop tests confirmed that the augmented model reduced prediction errors across all eight system states compared to the baseline ROM. The improvement in rotor angular velocity error alone was approximately 20 percent.
In closed-loop simulation at 11 m/s wind speed, the adaptive controller — labeled ENMPCaug — achieved 9 percent higher economic profit than the non-adaptive ENMPC baseline, alongside significantly lower pitch and torque travel. Switching to LiDAR-based wind speed estimates pushed that profit advantage to 30 percent. ENMPCaug with LiDAR input and only five solver iterations delivered 22 percent higher profit than the baseline running ten iterations, while requiring just 15 percent more computation time.
Limitations and Next Steps
The results come with important caveats. The current profit model accounts for tower base fatigue only — bearing wear, gearbox degradation, and blade fatigue are all absent, despite carrying real economic weight.
Validation has been conducted solely in simulation using the NREL 5 MW reference turbine. Field deployment and certification testing remain ahead. Real-time feasibility at the 100 ms sample rate also needs fuller demonstration, since the neural network adds CPU load at every control step.
What This Means in Practice
Correcting a reduced-order model’s blind spots with an offline-trained neural network can meaningfully improve both economic performance and actuator efficiency. The offline training approach is deliberate: it allows safety and performance to be verified before deployment, a prerequisite for industrial certification that online learning can’t easily satisfy. Profit gains are also sensitive to wind input quality — better wind information, such as LiDAR preview, amplifies the benefit of model adaptation considerably.
Future development will need to extend the fatigue model to additional turbine components, incorporate dynamic electricity market prices over longer time horizons, and explore online parameter tuning alongside the offline augmentation already demonstrated here.
