Model Predictive Control of a Wind Turbine
Authors
Haugaard, Thomas ; Martín Gómez, Álvaro ; Ajenjo De Torres, Oier
Term
4. term
Education
Publication year
2022
Submitted on
2022-06-01
Pages
153
Abstract
Wind turbines are becoming larger, which makes it important to use smarter control systems that both protect the structure and maximize electricity generation. This project tests Model Predictive Control (MPC), an advanced method that looks ahead to choose control actions while honoring limits, to reduce loads and optimize power. We built a non-linear model of a turbine that includes the motion of each blade so the pitch (blade angle) can be controlled independently. Because many internal quantities cannot be measured directly, we used an Unscented Kalman Filter (UKF) to estimate them, using data from the GH Bladed simulation tool. A large mismatch between our model and the Bladed outputs prevented acceptable results, so we temporarily used our own model outputs as "virtual measurements" by adding typical sensor noise. We then supplied the MPC with a prediction model obtained by successive linearization—approximating the system around its current operating point—and designed it to work in both partial- and full-load operation. Current results are not satisfactory because the linearized model’s states were treated incorrectly: they are defined relative to the operating point at each step. We expect performance to improve once this issue is fixed.
Vindmøller bliver større, og derfor er der brug for smartere styring, som både skåner strukturen og maksimerer elproduktionen. I dette projekt afprøves modelprædiktiv styring (MPC), en avanceret metode der ser et kort stykke frem i tiden for at vælge styregreb, samtidig med at grænser overholdes, så laster kan reduceres og effekt optimeres. Vi opbyggede en ikke-lineær model af en vindmølle, der inkluderer hver klings dynamik, så bladvinklen (pitch) kan styres uafhængigt for hver klinge. Fordi mange indre størrelser ikke kan måles direkte, brugte vi et Unscented Kalman-filter (UKF) til at estimere dem, baseret på data fra simuleringsværktøjet GH Bladed. En stor uoverensstemmelse mellem vores model og Bladed-output forhindrede acceptable resultater, så vi brugte midlertidigt vores egne modeloutput som "virtuelle målinger" ved at tilføje typisk sensorstøj. Derefter forsynede vi MPC med en forudsigelsesmodel opnået ved successiv linearisering—en trinvis approksimation omkring det aktuelle driftspunkt—og designede den til både dellast- og fuldlastdrift. De nuværende resultater er ikke tilfredsstillende, fordi de lineære modeltilstande blev behandlet forkert: de er defineret relativt til driftspunktet i hvert trin. Vi forventer forbedringer, når dette problem er løst.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords