Preconditioning wind speeds for standardised structural design

Abstract

The statistical treatment of historical wind data has a significant impact on the design wind load, standardisation of which requires characterisation of the wind climate across a large region. Improved spatial resolution is afforded by asymptotic extreme value distributions as they require a lower temporal data resolution. The Gumbel distribution is shown to be preferred, given the available South African data, according to the Akaike Information Criterion. However, its relative inflexibility means that estimates of small probability fractiles could be significantly biased, a difficulty illustrated by disagreement on which is more appropriate: fitting the Gumbel distribution to the wind speed versus to the wind-induced pressure (squared wind speed). By generalising this choice to the problem of finding the most appropriate real-valued exponent to raise the wind speeds by (i.e. not just 1 or 2), this paper aims to reduce this modelling-bias. As estimates of this exponent parameter from a single station would have low confidence given the typical quantity of data, a method of estimating the exponent that maximises the likelihood of observing the entire dataset is developed instead. The method is demonstrated using over 3500 annual gust measurements at 131 stations throughout South Africa, the most likely exponent is found to be 1.59 with confidence intervals sufficiently narrow to reject fitting to both the wind speed as well as to the squared wind speed. It is shown that using an exponent of 1.6 in deriving the design wind load for structural design would result in a substantial reduction in modelling-bias, providing a suitable baseline for future South African loading standards.