Comparison of Machine Learning-Based Wind Turbine Power Generation Prediction with Mathematical Modeling and Physical MATLAB/Simulink Approaches

Kurzfassung

This Master Thesis focuses in the comparison and evaluation of three approaches for the power prediction of the wind turbine OPUS 1 located in the Wivaldi wind farm in Krummendeich, Germany to identify their strengths and limitations. The first approach uses machine learning (ML) models. Seven commonly used ML models were initially tested, and three (Extreme Gradient Boosting (XGB), Random Forest (RF), and Gated Recurrent Units (GRU) were selected based on stable accuracy across training and testing datasets. These models achieve lower error metrics when input features are strongly correlated with the target variable. The second approach is a mathematical model, based on the turbines power coefficient curves derived from fundamental wind energy equations. It calculates power output from wind conditions using turbine characteristics. This approach provides stable and interpretable predictions with low computational demand. The third approach is a wind turbine simulation model implemented in MATLAB/Simulink, based on turbine specifications and steady-state aerodynamic relationships. It computes power output dynamically from wind conditions and turbine operational variables. Finally, the approaches are compared in terms of prediction accuracy, computational time, resource consumption, and complexity to identify the strengths and limitations. The best accuracy for most ML models was achieved with the features with a strong correlation with the target variable. For limited numbers of input features, Mathemat ical approaches perform with a higher accuracy than ML approaches, but they struggle to keep up with rapid changes in short periods of timere, flecting the inherent behavior of this simplified formulation.