Marefat, H. Ali (2025) Advancing machine learning for large eddy simulation of offshore wind farms. Doctoral (PhD) thesis, Memorial University of Newfoundland.
Full text not available from this repository.
Abstract
Subgrid-scale modelling in large eddy simulation remains an open problem in the field of turbulence modelling, particularly when dealing with complex ows like those in offshore wind farms. The core challenge in subgrid-scale modelling is accurately representing the effects of unresolved small-scale turbulent motions on the larger, resolved scales of the ow. Traditional models frequently rely on simplifying assumptions, such as isotropy, to characterize eddy viscosity. However, these assumptions often fall short in accurately capturing the complex, anisotropic, and multiscale nature of turbulence in real-world offshore conditions. Such limitations can lead to inaccuracies, including over- or under-dissipation of energy from the resolved scales, ultimately impacting the predictive accuracy and stability of large eddy simulations. Machine learning and data-driven approaches offer promising alternatives, focusing on developing robust, generalizable subgrid-scale models that enhance the scalability and performance of large eddy simulations without compromising computational efficiency. This thesis proposes an advanced machine-learning-facilitated approach to improve turbulence modelling within the large eddy simulation framework, with a focus on offshore wind farms. By incorporating machine learning techniques, this research addresses the inherent challenges of capturing the complex, multiscale turbulent ow behaviours typical in offshore environments. The work is centred on the development and assessment of machine-learning-based subgrid-scale models aimed at improving the predictability, performance, and scalability of turbulence models for offshore wind farm applications. To establish the machine-learning-based subgrid-scale model, an offshore wind farm was simulated using the actuator disk method, with simulated data compared against wind tunnel measurements before model training. Both standard a-priori and a-posteriori analyses were performed to evaluate the machine learning model's effectiveness comprehensively. The study began by leveraging a scale-adaptive large eddy simulation of ow past a sphere, which offers insights into turbulent wake dynamics and essential features of the turbulence energy cascade. Based on this analysis, a novel subgrid-scale model using a wavelet-assisted encoder-decoder architecture with skip connections was introduced to predict subgrid-scale stresses in the wake of a sphere using the Germano-Lilly framework. Results demonstrated that this model is particularly adept at capturing ow intermittency and preserving spatial ow information in the wake of a sphere at Reynolds numbers of Re = 103 and Re = 104. To enhance the scalability and generalizability of the wavelet-assisted encoderdecoder subgrid-scale model, the study undertook an in-depth examination of its interpretability. The efficacy of encoder-decoder models, particularly in turbulence modelling, depends significantly on their latent representations. As such, the nonlinearity of the encoder-decoder's latent space was scrutinized, revealing promising results in capturing the intermittency and chaotic nature of turbulent ows. Achieving scalability necessitates testing the model in both a-priori and a-posteriori setups, which led to the development of the Scale-Adaptive Machine-learning Subgrid-Scale (SAM-SGS) model for large eddy simulations of offshore wind farms. Building upon prior mathematical insights, this model captures subgrid-scale turbulent eddies by directly learning coherent enstrophy dynamics and energy cascades. Leveraging mixed subgrid-scale modelling principles, the SAM-SGS model employs the encoder-decoder architecture to represent structural stresses and uses eddy viscosity to model turbulence dissipation. The potential of the SAM-SGS model extends beyond its immediate accuracy in turbulent ow prediction. Its ability to dynamically adapt to varying ow conditions in offshore wind farms presents an opportunity for greater scalability and generalization in large-scale simulations. Through a-priori and a-posteriori analyses, it demonstrated a strong capability to capture essential turbulent features, underscoring its potential to enhance turbulence modelling in offshore wind energy technology. Notably, the SAM-SGS model accurately predicted subgrid-scale turbulence statistics under new initial conditions and even with different filter widths than those in the training data. This capability suggests robust generalizability to varying conditions, highlighting its promise for advancing offshore wind technology.
| Item Type: | Thesis (Doctoral (PhD)) |
|---|---|
| URI: | http://research.library.mun.ca/id/eprint/16894 |
| Item ID: | 16894 |
| Additional Information: | Includes bibliographical references (pages 147-176) -- Restricted until January 30, 2026 |
| Keywords: | turbulence modelling, machine learning, large eddy simulation, wind farm |
| Department(s): | Engineering and Applied Science, Faculty of Science, Faculty of > Mathematics and Statistics |
| Date: | February 2025 |
| Date Type: | Submission |
| Library of Congress Subject Heading: | Wind power plants; Machine learning; Turbulence; Computational fluid dynamics; Offshore wind power plants; Fluid dynamics--Mathematical models |
Actions (login required)
 |
View Item |
