Numerical modeling and machine learning for wind turbine aerodynamics and condition monitoring

• Main Author: Purohit, Shantanu
• Other Authors: Ng Yin Kwee
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Mechanical engineering::Fluid mechanics; Engineering::Mechanical engineering::Alternative, renewable energy sources
• Online Access: https://hdl.handle.net/10356/159299
• Institution: Nanyang Technological University

Wind energy is expected to play a vital role in transitioning from fossil-based energy sources to renewable energy. However, to make wind energy more competitive, it is essential to understand the underlying wind turbine aerodynamics, which plays a critical role in the performance of wind turbines. The study of wind turbine wakes, in particular, is crucial as it is responsible for power losses and fatigue loads on downstream turbines. As a result, wind turbine components have a higher chance of failure, and it becomes indispensable to identify the impending faults in wind turbine components. In the current thesis, three studies have been conducted that tackle these issues. In the first study, important light is shed on the rotor and wake aerodynamics of the (New) MEXICO rotor. Two turbulence flow modeling approaches have been compared viz. unsteady Reynolds-Averaged Navier-Stokes (uRANS) and Large-Eddy Simulation (LES) methods. The results of pressure and velocity components are compared against the experimental measurements. It is found that for the lower tip-speed ratio (TSR), uRANS shows faster wake recovery than the LES method. Moreover, the LES approach predicts flow separation point better than the uRANS approach. The second study proposes a novel framework based on Computational Fluid Dynamics (CFD) and Machine Learning (ML) to predict wake characteristics. The dataset used to train the ML models is generated using full-scale simulations of the NREL Phase VI wind turbine rotor for different wind speeds. The three ML models employed are Support Vector Regression (SVR), Artificial Neural Networks (ANN), and eXtreme Gradient Boosting (XGBoost). The predictions from these models are then compared against the existing analytical wake models and CFD results. It is found that wake models based on ML could predict the wake characteristics comparable to the CFD simulations and better than the existing wake models. Finally, in the third study, deep learning models based on Convolutional Neural Networks (CNN) and time-series models, such as Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), are proposed for condition monitoring of wind turbines. To this end, a Supervisory Control and Data Acquisition (SCADA) dataset of a wind farm is employed. The bearing temperature and generator temperature are the two variables that are monitored to predict the impending faults in wind turbine components. The hyperparameters of deep learning models are optimized using Bayesian optimization. It is found that the deep learning model based on CNN-GRU gives a better performance than CNN-LSTM and other benchmark models.