Dissertation Defense Announcement

Stochastic weather generators (SWGs) are statistically-based point-scale models of meteorological data that are driven by random number generators. Commonly taking observational data or lowresolution global climate model data as input, they are useful tools for generating many realizations of possible climate scenarios for use in impacts studies. This dissertation presents the stochastic harmonic autoregressive parametric (SHArP) weather generator. SHArP is based on previous SWGs but it generates air temperature values directly instead of prescribing and removing the mean and standard deviations in advance and generating temperature residuals. In addition, in both the precipitation process and the temperature process, SHArP includes nonstationarity due to oceanic modes of variability. During frontal passage, the precipitation-responsive autocorrelated transitions result in more realistic temperatures. The multisite generalization of SHArP presents a challenge due to an exponential increase in the number of noise coefficient matrices as the number of sites increases, but empirical orthogonal function analysis is applied to the precipitation patterns over the domain in order to reduce the number of noise coefficient matrices to a reasonable number. For multisite precipitation simulation, a trend due to climate change is added. Even though they are statistically-based, SWGs are limited in their ability to capture meteorological extremes, including dry and wet spells. The second-order Markovian probabilities of precipitation at a single site are modified using the method of large deviations. This mathematically-based method is shown to accurately modify the probabilities of precipitation to produce binary precipitation occurrence time series that are extreme yet statistically consistent with the input data without needing to "wait to get lucky" for those extreme events to occur in very long simulations.