Transcript Downscaling Methods EBx

Downscaling Techniques and Regional Climate
Modelling
PRECIS workshop
Tanzania Meteorological Agency, 29th June – 3rd July 2015
Objectives of the session
To review downscaling methods of obtaining fine-scale
climate information from global climate models (GCMs),
with an emphasis on regional climate models (RCMs).
Outline
1. Why downscaling?
2. How to downscale (downscaling
techniques)
 Statistical methods
 Dynamical methods
3. Suitability of downscaling techniques
Why downscaling?
Why downscaling?
GCM lack regional details due to coarse resolution
for many climate studies -> needs fine scale
information to be derived from GCM output.
• Smaller scale climate results from an interaction
between global climate and local physiographic
details and from unresolved motions and processes
• There is an increasing need to better understand the
processes that determine regional climate
• Impact assessors need regional detail to assess
vulnerability and possible adaptation strategies
Coarse spatial resolution is not the
only problem
At their typical resolutions,
GCMs can skilfully resolve
large weather systems (e.g
extra-tropical cyclones).
These systems have a
lifetime of several days.
Shorter time scales and
details of these systems
requires higher horizontal
resolutions
How to downscale?
Downscaling techniques
100-300km
Coarse atmospheric data (T, Q, winds, pressure etc)
• Statistical: based on statistical
relationship between large- and localscale, derived from present climate
fine scale value = F (large-scale variables)
• Dynamical: Numerical models at high
resolution over region of interest
limited area model (regional climate model)
high resolution AGCM
• Statistical/Dynamical method
50km-1km
Local surface data (T, rainfall, winds, Q etc)
Why do they work?
The largest fraction of the
energy of the atmospheric
motion is due to very large
scale systems. These
systems are the main drivers
of local scale weather.
from Cullen (2002), ECMWF
Statistical Downscaling
Categories of Statistical Downscaling
• Weather generators
• Stochastic methods designed to reproduce statistical
properties of local variables (mean, variance) and their
temporal structure. Include Markov chain, spell length
approach
• Transfer functions
• Methods based on fitting linear or non-linear regression
relationship between local variables and large scale
weather variables. Includes simple linear regression,
piecewise interpolation, artificial neural networks,
Generalised Linear Models etc.
• Weather typing
• Methods based on weather classification (cluster
analysis, self organising maps, etc) and analogue
method, local variables are related to the most relevant
large scale weather patterns
Assumptions made for Statistical
Downscaling
Relies on large-scale predictors for which Climate System Models
are most skilful:
• GCM skilful scale is assumed to be several grid lengths
• Dynamic variables (geopotential, wind, temperature)
• Tropospheric variables (away from the surface)
The transfer function must remain valid in the changed climate
(stationarity assumption)
• Hard to demonstrate
• Can be evaluated by comparison with other approaches
Predictors can be selected by numerical algorithms (e.g stepwise
regression)
• Importance of testing several predictors
• Uncertainties related to the choice of predictors
• Necessity to include predictors describing climate change
Advantages of Statistical Downscaling
• Computationally cheap
• The calibration of statistical models is usually done
only once and it can be applied to large sets of GCM
climate scenarios
• Produces local scale surface variables
• Many impact studies need “point value” surface
variables. These very high resolution are not easily
obtainable by climate modelling. Statistical methods
applicable to multi-site and multi-variable problems are
also available.
• Statistical downscaling packages available
• SDSM (weather generator), clim.pact (R package,
linear and analogue methods, not supported)
Regional Climate Modelling
What is a Regional Climate Model?
• Comprehensive physical high
resolution climate model that
covers a limited area of the globe
• Includes the atmosphere and
land surface components of the
climate system (at least)
• Contains representations of the
important processes within the
climate system
• e.g. clouds, radiation,
precipitation
One way nesting
• A RCM is a limited area model (LAM),
similar to those used in numerical
weather prediction (NWP), i.e. short term
weather forecasting
• LAMs are driven at the boundaries by
GCM or observed data
• Lateral (side) and bottom (sea surface)
• LAMs are highly dependent on their
boundary conditions and can not exist
without them
Lateral Boundary Conditions (I)
• LBCs = Meteorological boundary conditions
at the lateral (side) boundaries of the RCM
domain
• They constrain the RCM throughout its
simulation
• Provide the information the RCM needs from
outside its domain
LBC variables
LBC variables
• Data come from a GCM or reanalysis (quasiobserved lbcs)
• Lateral boundary condition variables
• Wind
• Temperature
• Water
LBC variables
• Pressure
• Aerosols
Lateral Boundary Conditions (II)
• Relaxation method (PRECIS)
• Large scale forcing merged with
internal solution over a lateral buffer
zone
• Large scale forcing of low wave
number components
• Important issues
• Spatial resolution of driving data
• Updating frequency of driving data
S. v.
RCM
interior
State variables
S. v.
• Spectral nesting
State variables
Sea Surface Boundary Conditions
• Two methods of supplying SST and sea ice:
• Using outputs from a coupled AOGCM
• Need good quality simulation of SST and sea ice in
model
• Necessary for future simulations
• Using observed values
• Useful for the present-day simulation.
• For future climate need add changes in SST and
ice from a coupled GCM to the observed values –
complicated
Simulation length
• Minimum period
• 10 years to reasonably study the mean climate
• Longer periods are better
• 30 years or more to study higher order statistics,
climate variability, extremes, etc
• Multi-annual mode of variability should also be
considered
Optimal choice of domain
• Continental scale (5000km x 5000km)
smaller domain do not allow the development of
mesoscale features, larger domain may loose the
consistency with large scale atmospheric flow (Jones et
al, QJRMS, 1995)
• Region of interest toward the middle of the
domain
• Buffer zone preferably placed on regions of
smooth orography.
Sources of errors in RCMs
• The RCM adds fine detail to the large-scale and shouldn’t
deviate from it, otherwise the one-way nesting approach
is not valid. The choice of the domain is important for
these reasons.
• Two sources of error:
• Large scale driving fields (external)
• Regional Model formulation (internal).
Added Value of RCMs
RCMs simulate
current climate more
realistically
Patterns of present-day winter
precipitation over Great Britain
Represent smaller islands
Projected changes in summer surface air temperature between present
day and the end of the 21st century.
Add physically consistent details to
climate change projections
Projected changes in winter precipitation between 1970s and 2080s.
Describe daily extremes more
accurately
Frequency of winter days over the Alps with different daily rainfall
thresholds.RCM and Obs aggregated at GCM scale
Resolve intense mesoscale systems
A tropical cyclone is evident in the RCM (right) but not in the GCM
Can be used to drive other models
A cyclone-like feature in the Bay of Bengal simulated by an
RCM and the resulting high water levels in the Bay simulated
by a coastal shelf model.
Suitability of Downscaling Techniques
Suitability of Regionalisation
Techniques
Method
Strengths
Weaknesses
Statistical
 High resolution
 Computationally
cheapscale (point value)
 Local
 Dependent on empirical relationships
derived for present-climate variability
 Not easily relocatable
 Few variables available
 Dependent on the availability of good quality obs
Regional climate
models
 High (very high)
resolution
 Can represent
extremes
 Physically based
 Many variables
 Easily relocatable
 Possible lack of two-way nesting
 Computationally expensive
 Dependent on driving model &
surface boundary conditions
 Have to parameterise across scales
(
)
Summary
• Downscaling techniques are used to add fine scale details
to GCM projections
• Statistical downscaling and dynamical downscaling are the
two most commonly used methods, each with its own
strengths and weaknesses.
• Impact studies for which downscaled scenarios are
needed and the design of the downscaling experiment will
determine whether to use one (or both) methods.
Questions
Criteria for Suitability of Downscaling Techniques
• Consistency at regional level with global projections
• Physical plausibility and realism
• Appropriateness of information for impact assessment
• Representativeness of the potential range of future climate change
• Accessibility for use in impact assessments