Direct Sampling for Extreme Events Generation and Spatial Variability
Enhancement of Weather Generators
Abstract
Weather generators based on resampling simulate new time series of
weather variables by reordering the observed values such that the
statistics of the simulated data are coherent with the observed ones.
These weather generators are fully data-driven and simple to implement,
do not rely on parametric distributions, and can reproduce the dynamics
among the weather variables under analysis. However, although the
simulated time series is new, the produced weather fields at arbitrary
timesteps are copies of the weather fields found in the training
dataset. Consequently, the spatial variability of simulations is
restricted. Furthermore, these weather generators cannot create weather
fields with out-of-sample extreme values because the scope of the
resampling method is constrained to the observed values. In this work,
we embedd the Direct Sampling algorithm — a data-driven method for
producing simulations — into resampling-based weather generators to
improve the spatial variability of the produced weather fields, and for
generating extreme weather fields. We increase the spatial variability
by applying Direct Sampling as a post-processing step on the weather
generator outputs. Furthermore, we produce out-of-sample extreme weather
fields using Direct Sampling in two ways: 1) applying quantile mappings
on the Direct Sampling simulations for a given return period, and 2)
using a set of control points jointly with Direct Sampling with values
informed by return period analysis. We validate our approach using
precipitation, temperature, and cloud cover weather-fields time series
datasets, for a region in northwest India. The results are analyzed
using a set of statistical and connectivity metrics.