Machine Learning-Assisted Dynamic Modelling of Wind Turbine Blades for Response Prediction

Abstract

The wind energy sector has seen significant growth in the last few decades in terms of the number of wind turbines constructed and the scale of these structures. Multimegawatt wind turbines on scales previously unimaginable are now routinely installed. Researchers are relentlessly pushing the boundaries of such structures, developing turbines with taller towers and longer blades to harness higher wind speeds and maximise energy capture. The increasing scale of these structures presents additional challenges in the form of reduced stiffness, increased loading, and larger deformations. In addition, to reduce the levelised cost of energy, all components of wind turbine systems are required to be light-weight and cost-effective but simultaneously satisfy performance and reliability requirements. These constraints require a more detailed analysis of the design, reliability, and fatigue of these structures.

The growing complexity of wind turbine designs, driven by the increasing scale of turbines, introduces significant challenges related to the stochastic nature of external loading, such as varying wind speeds and turbulence intensity. These environmental factors govern the dynamic loading on wind turbines, leading to variations in load distributions and uncertainty in achieving optimal performance. Since this uncertainty cannot be reduced, but must be quantified, traditional methods of analysis, which rely on static load envelopes, often fail to ensure accurate performance predictions across different installation sites. To address this challenge, a more dynamic, site-specific approach is needed, where the unique loading conditions at each installation location should be used for performance analysis. However, the high computational cost of such simulations, which involves modelling complex aerodynamic loads and deformations, makes this approach computationally challenging for large-scale use. This thesis aims to alleviate this issue by developing computationally efficient surrogate models to emulate the dynamic response of wind turbine.

In this thesis, a multi-body dynamic (MBD) model of a wind turbine is developed to model the onshore variant of the International Energy Agency�s (IEA) 15-MW wind turbine. The high computational cost of performing dynamic analysis of wind turbines subjected to turbulent wind inflow via the MBD model stems from two major factors: computation of aerodynamic loads using Blade Element Momentum (BEM) theory and computing the deformation response using a time stepping-based numerical integration algorithm. To address this, an Artificial Neural Network (ANN) model is developed using a data-driven and physics-informed approach to predict the aerodynamic loads. This model bypasses the computationally expensive BEM theory for quick estimation of aerodynamic loads. The resulting ANN-based surrogate model could predict the angle of attack at investigated blade nodes with the maximum Mean Absolute Error (MAE) of the order of $10^{-2}$. Building up on this, a Long Short-Term Memory (LSTM) model is developed to predict the blade deformation response when subjected to time-varying aerodynamic loads. These models combined can substantially reduce the computational cost of performing dynamic analysis on wind turbine blades. A combined wrapper and filter-based approach to identify the most informative features and a learning rate scheduling approach to optimize the model training are developed to establish an efficient workflow for developing such models. In addition, these models offer the potential to serve as virtual sensors to digitally measure the aerodynamic loads and blade deformations. The insights from the feature selection approach can also guide sensor placement strategies. Further, an LSTM model is developed to predict the response of blade degrees of freedom (DOFs) when subjected to turbulent spatio-temporal wind fields. The Principle Component Analysis (PCA) and Discrete Cosine Transform (DCT) feature extraction methods are combined with the recursive feature addition (RFA) method to reduce the high dimensionality of the wind field and arrive at optimized input features for an LSTM model. The LSTM model developed using these features results in high-accuracy prediction at individual blade DOFs. In the validation stage, the LSTM model prediction of blade response achieved high accuracy with maximum Normalised Root Mean Square Error (NRMSE) attaining values of the order of $10^{-1}$. This LSTM model can predict the blade response without explicit knowledge of blade properties. This method presents a particularly promising approach for analysing the performance of commercial wind turbine models where the blade properties can not be shared due to intellectual property concerns. Overall, the models presented in this thesis can help in enabling quick estimate of aerodynamic loads and blade response, which can aid in efficient site-specific analysis, quick fatigue-life estimation and faster design exploration.