Modelling Power From Wind Turbines - A Bayesian Approach
Translated title
Modellering af Effekt Fra Vindmøller - En Bayesiansk Tilgang
Authors
Villadsen, Jonas ; Kallehauge, Tobias
Term
4. semester
Education
Publication year
2020
Submitted on
2020-06-03
Pages
95
Abstract
Vind-til-effekt (W2P) modeller oversætter vindforhold til den elektriske effekt, en turbine kan levere. I dag er mange W2P-modeller overvejende datadrevne og tager kun begrænset hensyn til de fysiske processer. Denne afhandling undersøger, hvordan man kan bygge en fysikbaseret W2P-model, der bruger forudgående statistisk viden om vindhastigheder på et bestemt sted. Den statistiske opstilling følger en metode og designfilosofi udviklet til projektet, kaldet Approximate Bayesian Total Error Analysis. Modelparametre estimeres med Markov Chain Monte Carlo (MCMC), en samplingteknik. To modeller, AW2P og NTF-AW2P, udvikles og testes mod etablerede W2P-modeller. De to nye modeller opfører sig ens, men på tværs af flere fejlkriterier klarer de sig på niveau med eller dårligere end eksisterende tilgange. Det peger på, at der er behov for yderligere udvikling for at forbedre de fysikbaserede modeller.
Wind-to-Power (W2P) models translate wind conditions into the electrical power a turbine can produce. Today, many W2P models are mostly data-driven, with limited attention to the underlying physics. This thesis explores how to build a physics-based W2P model that uses prior statistical knowledge about wind speeds at a specific site. The statistical setup follows a method and design philosophy developed for the project, called Approximate Bayesian Total Error Analysis. Model parameters are estimated with Markov Chain Monte Carlo (MCMC), a sampling technique. Two models, AW2P and NTF-AW2P, are developed and tested against established W2P models. The two new models behave similarly to each other, but across several error metrics they perform on par with, or worse than, existing approaches. This suggests that further development is needed to improve the physics-based models.
[This abstract was generated with the help of AI]
Keywords
Documents