Simulation of extreme weather events by a stochastic weather generator

Abstract

A stochastic weather generator is a model capable of generating daily weather patterns that are statistically similar to the observed patterns. Weather generators are commonly used in climate change studies as a computationally inexpensive tool to generate high resolution climate change scenarios based on the output from global climate models. Considering that the frequency and the magnitude of extreme weather events are likely to increase under climate change, there is a growing need to investigate how well weather extremes are simulated by weather generators. The aim of this study was to test the skill of the LARS-WG stochastic weather generator to simulate extreme weather events at 20 locations with diverse climates. Yearly maxima of daily precipitation, maximum temperature and length of heat waves, and their 10 and 20 yr return values were compared for observed and synthetic data by fitting the generalized extreme value distribution and computing confidence intervals. Means of yearly maxima and return values of daily synthetic precipitation were within the 95% confidence intervals (CI 95 ) of observed data for all sites. Daily maximum temperature extremes were reproduced less accurately. Although the root mean squared error (RMSE) calculated for the means of maxima of maximum temperature was <1°C, synthetic means for approximately half of the sites were outside the CI 95 for observed values. This indicates that the assumption used in LARS-WG, that daily temperature could be approximated by the normal distribution, is inadequate. Means of yearly maxima for length of heat waves and 10 and 20 yr return values were within the CI 95 for all sites except 3. For those sites where LARS-WG performance was inadequate, daily maximum temperature was not normally distributed but was skewed.