A simulative analysis of physically-motivated Koopman lifted states for wind farm model predictive control

This paper investigates how the selection of data points from a simulated wind field, combined with nonlinear lifting functions, influences wind turbine power predictions. The objective is to develop accurate linear models within the Koopman operator framework suitable for real-time Model Predictive Control. Using Dynamic Mode Decomposition, we compare three data configurations from a high-fidelity simulation: a dense wind field with 50,000 points, a dataset restricted to turbine locations, and a sparse dataset including two additional streamwise positions. Results demonstrate that models using data from well-placed sensors can outperform those using dense wind field data. Furthermore, while turbine-only data initially yields lower accuracy, we show that physically motivated Koopman lifting functions effectively compensate for missing spatial information. By incorporating these additional prediction features, turbine-only models achieve prediction accuracies comparable to those utilizing the complete wind field data, which relies on information density unavailable for real-world turbine control.

DDC Class

621.3: Electrical Engineering, Electronic Engineering

530: Physics

Lizenz

https://creativecommons.org/licenses/by/4.0/

Publication version

publishedVersion

Name

Dittmer_2026_J._Phys.__Conf._Ser._3224_032085.pdf

Size

1.37 MB

Format

Adobe PDF

Options

A simulative analysis of physically-motivated Koopman lifted states for wind farm model predictive control