||Henrik Asmuth <email@example.com>|
||2022-09-01 – 2023-03-01|
The complexity of Northern European wind conditions and new grid requirements for wind farm control challenge the state of industrial wind farm modelling. High-fidelity large-eddy simulation (LES) is considered the most accurate tool for such applications. However, the large computational demand limits the use of conventional LES to fundamental academic studies. The lattice Boltzmann method (LBM) states an alternative to classical LES models and offers significantly higher computational efficiency, particularly when using tailor-made GPU frameworks. Already now, GPU-based LBM-LES is replacing lower fidelity flow models in various other industries.
This project aims to improve the prediction of wind turbine power and loads in typical Swedish conditions (heterogeneous patchy forest and a mildly complex orography) through the direct and indirect use of LBM-LES.
Firstly, remaining gaps in the wind-energy-specific modelling capabilities of LBM models are to be closed. As for typical Northern European conditions this refers to the effects of stratification, forest canopies and complex terrain, all of which are typically badly captured by industry-standard engineering models. The resulting LBM framework will finally allow for high-fidelity modelling that captures these effects with run-times that are acceptable in the industrial practice. This alone holds the potential for notable uncertainty reductions in power and load estimations.
The second objective is to speed-up power and load predictions of wind turbines even further by establishing new surrogate models based on the LES results. To this end, different deep learning techniques (among others, convolutional neural networks) will be investigated. These surrogate models are trained with LES-generated data of the flow field and turbine response. Such models are potential replacements for existing semi-analytical engineering models that bring about notable improvements in accuracy, while having almost negligible run time. The latter has been confirmed by initial investigations of the approach. On the other hand, initial studies also show that large amounts of data (hence, LES runs) is required in order to train the model for all relevant operational conditions. Generating the training data and training the model thus hinges on efficient high-fidelity models like LBM-LES, and on the availability of large modern GPUs.