Validation of a stochastic temperature generator focusing on extremes, and an example of use for climate change

摘要

ABSTRACT: We present a stochastic seasonal functional heteroscedastic auto-regressive model developed to simulate daily (minimum, maximum, or mean) temperature time series coherent with observed time series and designed to reliably reproduce extreme values through a careful study of extremes and the fact that the tails of the distribution are bounded. The model was first validated using different daily minimum and maximum weather station time series over Eurasia and the US in different climatic regions. The model was able to produce coherent results both for the bulk of the distribution and for its extremes and was able to produce higher or lower extreme values than observed. A possible use in the climate change context was then tested. We fit the model over the first part of a long temperature time series and then used it to simulate a large number of possible trajectories for the second part when temperature increased. Two approaches were tested to do so, one based on a simple mean change in mean and variance and the other in considering the full seasonalities and trends estimated over the observed second part of the time series. Both approaches yielded good results, both for the bulk and for the extremes of the temperature distribution over the second part of the period. However, the second approach allowed us to take interannual variability changes into account, which leads to more realistic results when this occurs. Our results support the use of this tool for statistical downscaling, enabling the reliable reproduction of temperature extremes.