Multisite Generalization of the SHArP Weather Generator

Generalization of point-scale stochastic weather generators to simultaneously produce output at multiple sites provides more powerful support for hydrology and climate change impact studies. Generalization preserves the statistical properties of each individual site while maintaining proper spatial correlation over the domain. Here, generalization of the daily precipitation and temperature components of the stochastic harmonic autoregressive parametric (SHArP) weather generator is presented. The generalization process for temperature involves conversion of vector time series to matrix time series that capture between-site covariances of maximum and minimum daily temperature. Between-site temperature covariances depend on spatial precipitation-occurrence patterns (POPs), of which there are up to 2M for M sites. To dramatically reduce the number of covariance matrices that drive temperature, multisite SHArP uses empirical orthogonal function analysis to categorize the POPs and harmonic smoothing to reduce the number of parameters describing the temporal evolution (annual cycle) of the elements in the covariance matrices. By modeling precipitation-regime-specific residuals, the model is shown to capture statistically significant spatial and temporal contrasts in observed temperature covariance. For precipitation simulation, we extend existing techniques by adding a trend term to the occurrence and amount parameters. Multisite generalization of the framework is illustrated by simulating stochastic historical and future temperature and precipitation across complex terrain over northern Utah on the basis of historical station observations and historical and future statistically downscaled climate model output.