development of a downscaling prediction system liqiang sun international research institute for...

Development of a downscaling prediction system

Liqiang SunInternational Research Institute for Climate and Society (IRI)

• Downscaling – the translation of a forecast to a spatial or temporal resolution that is finer than that of the original forecast.

Definition

coarse resolution fine resolution

Definitioncoarse resolution

fine resolution

statistical model

dynamical model

GCM forecasts are area-averages, and may not represent a scale at which a forecast is to be used. A typical GCM grid is about 60,000 km2. Six million hectares is a rather large farm!

A part of the world according to a GCM...

red: land white: water

1 box ~ 90,000 km2

Motivation

Climate can vary dramatically over short distances, especially in the context of precipitation and wind speeds.

Small-scale affects (such as topography) important to local climate could be poorly represented in GCMs

Annual mean precipitation over western North America.

Motivation

• Teleconnection patterns can have detailed spatial structure, at resolutions too fine for GCMs.

Motivation

GCMs have approx. 20 min. timestep. But GCMs do not simulate sub-monthly weather phenomena well.

Climate downscaling (spatially and temporally)

Enhancing the scale and relevance of seasonal climate forecasts and creating information to better support decisions

Downscaling TechniquesDynamical Downscaing - Localized climate information is generated using high resolution regional climate models (RCMs), driven by low resolution global climate models (GCMs), or using a variable resolution global model in which the highest resolution is over an area of interest.

Statistical downscaling involves relating the large scale climate state to target variables using a transfer function (e.g., regression).

observed large scale climate

Transfer function(statistical model)

observed small scale climate

predicted large scale climate

predicted small scale climate

Strengt/weakness of SD and DD (1)

from: Wilby and Dawson, 2004: Using SDSM Version 3.1-A decision support tool for the assessment of regional climate change impacts.

Strengt/weakness of SD and DD (2)

Statistical downscaling is feasible for many particular regions, and is the appropriate baseline upon to measure the success of the much more costly dynamical approach

Development of Regional Climate Prediction System

choosing GCM forecasts that have good skill over the region of interest Identifying a group of regional climate models for downscaling Determining model resolution and domain sizeCustomization of regional climate modelsObservationInitializationPrograms to nest regional models within GCMsEnsemble runs of retrospective forecast – forecast skillStatistical post processing of model output to correct model biasForecast productForecast verification

CLIMATE DYNAMICAL DOWNSCALING FORECAST SYSTEM FOR NORDESTE

PERSISTED GLOBAL SST ANOMALIES

ECHAM4.5 AGCM (T42)NCAR CAMS

AGCM INITIAL CONDITIONS

UPDATED ENSEMBLES (10+)WITH OBSERVED SSTs

Persisted SSTA ensembles 1 Mo. lead

Predicted SSTA ensembles 1-4 Mo. lead

10

24

PostProcessing

RSM97 (60km)RAMS (40km)

CPT

HISTORICAL DATA•Extended Simulations•Observations

PREDICTED SST ANOMALIES

Tropical Pacific Ocean(LDEO Dynamical Model)(NCEP Dynamical Model) (NCEP Statistical CA Model)Tropical Altantic Ocean(CPTEC Statistical CCA Model)Tropical Indian Ocean(IRI Statistical CCA Model)Extratropical Oceans(Damped Persistence)

IRI FUNCEME

Sun et al. (2006)

Model Resolution

1. The horizontal and vertical resolutions should be fine enough to capture the scales of forcings of interest (e.g., Spatial characteristics of the land surface forcing)

Model Topography

300KM

10KM

50KM

Model Resolution (Cont.)

2. Availability of model dynamical and physical parameterizations. All the parameterization schemes are based on a spectral gap between the scales being parameterized and those being resolved on the grid. Therefore, all the model parameterization schemes are model resolution dependent.

For example, cumulus parameterization

Grid Spacing (KM)

5 20 40

______|_______|_____________|__________

explicit ??? hybrid GCMs

Model Resolution (Cont.)

3. The ratio of driving data versus RCM horizontal resolution is in the range of 3-8 ( for traditional one-way nesting approach). If the downscaling grid space ratio is too large, multiple nesting is sometimes used.

80km 300km

20km

Domain Size

• The area of interest is as far as possible from the lateral buffer zone.

• Model domain should encompass all regions that include forcings and circulations which directly affect climate over the area of interest.

• It is preferable to place the lateral boundaries over the ocean rather than land, especially not over areas of complex topography.

• It is preferable not to place the lateral boundaries over the areas with strong convection.

• Internal variability usually increases with domain size• Computational Limitation

Domain Size (Cont.)

Sun et al. (2005)

Forecastarea

Physics adequacy

All the model parameterization schemes are model resolution dependent because the parameterization schemes are based on a spectral gap between the scales being parameterized and those being resolved on the grid.

Network of rainfall stations available for the indicated datasets and dates

Horizontal and vertical interpolation errors

Horizontal: the grid-point spacing and the map projections are different between the RCM and the GCM.

Vertical: the grid-point spacing and the coordinates are different between the RCM and GCM.

Particularly the difference between the topographic field due to the different resolutions imply that extrapolations below the surface of the driving GCM have to be performed.

Spin-up

• Spin-up period is the time that model takes to achieve its climate equilibrium

• Spin-up time varies depending on the domain size, season, circulation strength, and surface conditions.

Atmospheric component - daysLand surface component:

top layer 0.1m - weeksroot zone 1.0m - monthsdeeper soil >1.0m - years

Ocean component:upper ocean 500m - tens yearsdeep ocean >500m - hundreds years

Land Initialization

An offline land model should be used to generate the land initial conditions for the regional climate model, or

use of reanalysis data for land initialization. The resolution of the reanalysis data should be the same as or similar to that of the regional climate model. The reanalysis data should be statistically

corrected.

1) calculate the standard anomaly,

2) perform statistically correction

3) corrected anomaly added to

the model climatology

Update frequency of the driving data

1. This issue has to do with the temporal resolution of the dataset used to drive the nested RCM. As a rule of thumb, the updated period should be smaller than one quarter of the ratio of the length scale of the phase speed of the meteorological phenomena that we want to get correctly in the model domain. For instance, A typical synoptic system having a horizontal size of 1000 km and a phase speed of 50 km/h would require an updating frequency of at least 5 h.

Diurnal variation is important for the tropics, it would require an updating frequency of at least 6 h.

2. Increasing the updating frequency will also introduce nesting noise. The nesting frequency of 3-6 h is mostly used.

Quality of the driving data

This issue has basic implications in the concept of nested RCMs because even with a perfect model and perfect nesting scheme, the quality of the driving data used is very important.

In case of a GCM driving an RCM, if the GCM large-scale circulation prediction is wrong, good results cannot be expected from the RCM. In other words, “garbage in”

“garbage out”.

Bias correction for the driving GCM data is recommended

Temporal anomaly correlations between the observed and the model ensemble mean rainfall

Geographical distributions of RPSS (%) for the hindcasts averaged over the period of 1971-2000

Model Output Statistics (MOS)

Jjj

Address relevant scales and quantities –climate variables that are both relevant and predictable

Precipitation

Temperature

Extreme events

Onset of rainy season

Dry spell & wet spell

Tropical cyclones

…


choosing GCM forecasts that have good skill over the region of interest Identifying a group of regional climate models for downscaling Determining model resolution and domain sizeCustomization of regional climate modelsObservationInitializationPrograms to nest regional models within GCMs

• Ensemble runs of retrospective forecast – forecast skillStatistical post processing of model output to correct model biasForecast productForecast verification

Real-Time Forecast Validation


• choosing GCM forecasts that have good skill over the region of interest

• Identifying a group of regional climate models for downscaling

• Determining model resolution and domain size• Customization of regional climate models• Observation• Initialization• Programs to nest regional models within GCMs• Ensemble runs of retrospective forecast – forecast skill• Statistical post processing of model output to correct model

bias• Forecast product• Forecast verification

development of a downscaling prediction system liqiang sun international research institute for...

Documents

climate downscaling

regional models

local climate

model resolution

large scale climate

original forecast

highest resolution

statistical downscaling