A spatio-temporal dynamic regression model for extreme wind speeds

This paper presents a hierarchical regression type model for analyzing the dependency of sample extremes on time, space and a covariate effect. The model is based on the assumption that the observations follow independently a generalized extreme value distribution given location, scale and shape parameters. Then a multivariate spatial process is considered to accommodate the association and spatial correlation in the distribution parameters. The mean of the process incorporates the underlying dynamics which is elaborated on the lower stage of hierarchy. Finally, three spatio-temporal dynamic linear models drive independently this mean function to take the variations in the parameters separately into account. In a Bayesian setting, the model structure leads to parallel implementation of the Markov chain Monte Carlo algorithm in a sense that it is less time consuming. Our methodology is applied to the monthly maxima of wind speed with temperature as a covariate for which the relationship is expressed in terms of a penalized spline regression model. The comparison of the proposed model with several simpler ones suggests considerable improvements in wind speed analysis.