| The accurate estimation of extremes of wind speeds is important for the optimum design of buildings, bridges and other structures. Most design codes specify design wind speeds for typical return periods of 30, 50 and 100 years. Currently, the design wind speed estimation in North American codes is based on relatively simple statistical analysis of wind, speed data collected at a particular site. For example, the design wind speeds specified in the National Building Code of Canada have been derived from the Gumbel distribution fitted to annual maximum wind speed data collected up to the early 1990's. A major drawback of the traditional Gumbel method is that it discards a large body of wind speed data that are collected on hourly basis, which deteriorates the accuracy of estimation due to large sampling and modelling errors. This has motivated the development of new concepts and methods for statistical analysis of wind speed data.; In recent years, advanced methods have been developed in the field of mathematical statistics to reduce both sampling and model uncertainties associated with extreme quantile estimates.; The objective of the thesis is to explore the use of advanced concepts and methods of extreme value theory for updating the design wind speed estimates specified in the Canadian and American design codes. New concepts and methods include enlarging the data set beyond annual maxima, improve the plotting position, and pooling the data from different sites to create superstations or regions.; The thesis presents a comprehensive investigation of the statistical analysis of wind speed data on the basis of at-site and the regional pooling of data. The at-site analyses of the Canadian and American data are carried out using the traditional Gumbel, modified Gumbel, independent storm method, peaks over threshold and annually r-largest order statistics (r-LOS). A comparative assessment of these methods provides valuable insight about the effects of assumptions associated with different methods on the accuracy of quantile estimation. The thesis concludes that the r-LOS is the most versatile method, as it works with the Generalized Extreme Value distribution. The r-LOS method is an effective alternative for improving the estimation of extreme design speed.; The concept of regionalization is useful in reducing the sampling error associated with the estimation of extreme wind, especially when at-site records are relatively short. In this method, data from several sites are combined to estimate the wind speed distribution at a particular site of interest. A key step of this method is to identify statistically similar sites that can constitute a region. To improve the regionalization analysis, the thesis proposes an MCD-based discordancy measure and bootstrap-based heterogeneity measure, and validates them through Monte Carlo simulations. The proposed regionalization method is applied to both Canadian and American data and results are compared with existing code provisions. The results indicate that the proposed regional model can improve the wind speed analysis and provide more reliable statistical estimates.; A comprehensive investigation of at-site and regional models for the statistical analysis of wind speed data and the practical applications illustrated in the thesis are expected to advance the frontier of wind engineering and lead to more efficient and safe design of structures. |
