Characterizing weather and climate variability for precipitation: A data-based stochastic modeling framework

Abstract:

Precipitation is a notoriously noisy variable at spatial scales from rain-gauges to GCM grid cells, partially due to the complexity of occurrence processes. To properly characterize variability of precipitation at so-called "weather" versus "climate" time scales, we demonstrate a probabilistic method for representing daily precipitation within a stationary-climate ("weather only") framework with applications for constraining unrealistic precipitation variability in climate models. This method accurately recreates the annual seasonal cycle and the daily variability of precipitation conditioned on the previous m days of data in a generalization of an autoregressive process optimal for zero-inflated empirical probability distributions.

Monte Carlo simulations from these models create "weather" distributions for a given location at any spatial scale, which can then be compared to observations to determine the magnitude of "climate" variability (i.e., variability not well-represented by our stationary models). Weather distributions have applications in characterizing forecast confidence intervals, and climate/weather variance ratios allow us to calculate potential predictability for precipitation. Techniques for empirical model selection and calibration are highlighted, as are common assumptions about precipitation stationarity and variability. As an application, the variability of precipitation at weather/climate time scales is compared for observational data and a collection of historical runs from Global Climate Models.