Machine learning for wind flow modelling

: using grid-based neural networks to capture wind flow changes over terrain

Abstract

Modelling the wind flow over terrain is a key element of wind resource assessments within the wind energy industry. Existing flow modelling methods range from fast, low fidelity analytical models to time-consuming and computationally expensive high-fidelity Computational Fluid Dynamics software. Machine learning offers the potential for high-fidelity yet fast-running surrogate flow models. This project created surrogate wind flow models using machine learning which aim to achieve the accuracy of the industry-standard WAsP software for wind over terrain. WAsP is split into three components: orographic (elevation-induced) speedup; orographic turn; and roughness speedup. Hence, surrogate models were developed for each. While initial tests with Convolutional Neural Networks were unsuccessful, a Grid Neural Network approach was developed as part of this work, which takes in sub-grids of each input (terrain) map and output (speed or direction change) map, and uses data points from these sub-grids as the inputs and outputs of a Deep Neural Network. Models using this novel architecture proved to be trainable with a relatively small set of data, with the optimal sub-grid sizes discovered providing an understandable measure of the radius of influence of each input variable on the corresponding output. Using Grid Neural Networks, surrogate models were created to predict the orographic speedup and turn, and roughness speedup, at heights of 10m and 100m above ground level. The predictions for orographic speedup and turn for multiple sites and wind directions correlated well to the WAsP data at both heights. The surrogate model predictions for the roughness speedup at 10m above ground level were also a close match to the WAsP values, but at 100m above ground level the roughness speedup predictions were inaccurate for some sites. Future work could combine the surrogate models at separate heights, and incorporate the separate sub-model predictions into full wind resource maps.

Date of Award	22 Mar 2022
Original language	English
Awarding Institution	University of Bristol
Supervisor	Daniel J Poole (Supervisor) & Paul W Harper (Supervisor)