DSensemble: Downscale ensemble runs

In metno/esd.test: Climate analysis and empirical-statistical downscaling (ESD) package for monthly and daily data.

Description

Downscales an ensemble of climate model runs, e.g. CMIP5, taking the results to be seasonal climate statistics. For temperature, the result hold the seasonal mean and standard deviation, whereas for precipitation, the results hold the wet-day mean, the wet-day frequency, and the wet/dry-spell statistics. The call assumes that netCDF files containing the climate model ensemble runs are stores in a file structure, linked to the path argument and the rcp argument.

These methods are based on DS, and DSensemble is designed to make a number of checks and evaluations in addition to performing the DS on an ensemble of models. It is based on a similar philosophy as the old R-package 'clim.pact', but there is a new default way of handling the predictors. In order to attempt to ensure a degree of consistency between the downscaled results and those of the GCMs, a fist covariate is introduced before the principal components (PCs) describing the EOFs. The argument area.mean.expl=TRUE will take the time series describing area mean value for the selected predictor domain as the first covariate, followed by the PCs. These are then used in the regression analysis.

The argument non.stationarity.check is used to conduct an additional test, taking the GCM results as 'pseudo-reality' where the predictand is replaced by GCM results interpolated to the same location as the provided predictand. The time series with interpolated values are then used as predictor in calibrating the model, and used to predict future values. This set of prediction is then compared with the interpolated value itself to see if the dependency between the large and small scales in the model world is non-stationary.

Other chekch include cross-validation (crossval) and diagnostics comparing the sample of ensemble results with the observations: number of observations outside the predicted 90-percent conf. int and comparing trends for the past.

Usage

DSensemble(y,...)
DSensemble.default(y,path='CMIP5.monthly/',rcp='rcp45',...) 
DSensemble.t2m(y,plot=TRUE,path="CMIP5.monthly/",
                           predictor="ERA40_t2m_mon.nc",
                           rcp="rcp45",biascorrect=FALSE,
                           area.mean.expl=FALSE,
                           non.stationarity.check=FALSE,
                           eofs=1:6,lon=c(-15,15),lat=c(-10,10),
                           select=NULL,FUN="mean",FUNX="mean",
                           pattern="tas_Amon_ens_",verbose=FALSE)
DSensemble.precip(y,plot=TRUE,path="CMIP5.monthly/",
                              rcp="rcp45",biascorrect=FALSE,
                              predictor="ERA40_pr_mon.nc",
                              area.mean.expl=FALSE,
                              non.stationarity.check=FALSE,
                              eofs=1:6,lon=c(-15,15),lat=c(-10,10),
                              select=NULL,FUN="exceedance",
                              FUNX="sum",threshold=1,
                              pattern="pr_Amon_ens_",verbose=FALSE)
DSensemble.mu(y,plot=TRUE,path="CMIP5.monthly/",
              rcp="rcp45",biascorrect=FALSE,
              predictor="ERA40_t2m_mon.nc",
              non.stationarity.check=FALSE,
              eofs=1:16,lon=c(-30,20),lat=c(-20,10),
              select=NULL,FUN="wetmean",
              FUNX="C.C.eq",threshold=1,
              pattern="tas_Amon_ens_",verbose=FALSE)

Arguments

y	A station object.
plot	Plot intermediate results if TRUE.
path	The path where the GCM results are stored.
rcp	Which (RCP) scenario
area.mean.expl	When TRUE, subtract the area mean for the domain and use as a the first co-variate before the PCs from the EOF analysis.

Value

A 'dsensembele' object - a list object holding DS-results.

Author(s)

R.E. Benestad and A. Mezghani

Examples

y <- station.metnod("Oslo - Blindern")
rcp4.5 <- DSensemble(subset(y,is=1),plot=TRUE)
plot(rcp4.5)

# Evaluation: (1) combare the past trend with downscaled trends for same
# interval by ranking and by fitting a Gaussian to the model ensemble;
# (2) estimate the probabilty for the counts outside the 90
# percent confidence interval according to a binomial distribution.

dstest <- diagnose(rcp4.5)
plot(dstest)

metno/esd.test documentation built on May 22, 2019, 7:49 p.m.