Multi-Scale Statistical Evaluation of CMIP5 Predictions of Extreme Precipitation Events

In view of risk to critical infrastructure, many decision makers are concerned about extreme precipitation potential at specific locations; yet modeling of precipitation at local, regional and global scale remains uncertain and in many regions, problematic. We examine both physical and statistical strategies for improvement in predictions at global and regional scale. Various probability distribution functions are fit to extreme precipitation data such as National Climatic Data Center (NCDC) observational data, National Aeronautics and Space Administration (NASA) satellite observations and location-specific point gauge measurements in order to determine which statistical method gives the best predictive values for past observed extremes for a given location at various resolutions. While the Log Pearson III distribution shows better return-time extreme precipitation predictive capability than the Gumbel, or Type I Extreme Value Distribution for specific locations at coarse resolution (2x2.5 degree), physically-based regionally- and seasonally-informed methods incorporating both time and space parameters may better suit local and short term intensity predictions. Methods applied to Coupled Model Intercomparison Project Phase 5 (CMIP5) climate model historical data determine model capability to capture not only precipitation averages or maxima but also multi-decadal probability distributions. Uncertainty quantification of global and dynamically-downscaled Representative Concentration Pathway (RCP) 4.5 and 8.5 scenario outputs derived from these methods serve accordingly to determine regional and local risk predictions associated with climate change.