Posters

Abstract

This study introduces a novel hybrid model aimed at optimizing wind turbine control systems by integrating physics-based modeling with data-driven machine learning techniques. Traditional wind turbine modeling approaches either rely solely on physical equations, which are prone to inaccuracies due to system uncertainties, or on machine learning models, which often lack generalization to unseen data. To overcome these challenges, a physics-based, data-augmented model was developed. This hybrid model combines the strengths of both approaches, leveraging physical parameters such as rotor and generator inertia while incorporating operational data to train a neural network. The model predicts the effective wind speed across the turbine's rotor, which in turn improves the estimation of key variables such as aerodynamic torque and rotor speed. The wind turbine selected for this study is the 5 MW National Renewable Energy Laboratory (NREL) wind turbine. Data for the machine learning component of the hybrid model was generated using the OpenFAST software, in conjunction with the NREL Reference OpenSource Controller (ROSCO). OpenFAST provides high-fidelity, multi-physics simulations of wind turbine dynamics, while ROSCO serves as a controller designed to closely resemble industrial controllers, ensuring the dataset reflects realistic operating conditions. However, to enhance computational efficiency, the hybrid model employed lower-fidelity equations that are simpler to interpret and integrate with the data-driven component while still maintaining sufficient accuracy to capture the turbine's key dynamics. The model was trained on simulated data and tested against unseen wind profiles, demonstrating acceptable accuracy in predicting system states. This hybrid approach enables more precise control actions, contributing to smoother and more efficient wind turbine operation, and overall improved performance in real-world applications.