Transcript How Good Weather Forecasts Go Bad

Why Good Forecasts Go Bad
Dr Jonathan Fairman
21 April 2016
Presentation by Prof. Dave Schultz
Early meteorology was not a science.
“Whatever may be the
progress of sciences,
NEVER will observers
who are trust-worthy, and
careful of their reputation,
venture to foretell the
state of the weather.”
François Arago (1846)
Director of Meteorology Department: 1854–1865
13 observing stations around
Britain and 5 on the Continent
Table for the newspapers
Coastal signals for mariners
Daily 48-h forecasts of wind
speed, direction, and coming
weather for five regions in British
Isles
Weather Book (1862)
Robert Fitzroy
Criticism of the Forecasts
Scientific “gentlemen” of the
Royal Society belittled weather
observers and Fitzroy
“so inaccurate and haphazard
a character, as not to be of any
true scientific value”
Robert Fitzroy
Numerical weather prediction as a
process is fairly well understood
So, meteorologists should make
perfect weather forecasts, right?
However,
occasionally
things go
wrong
Rodwell et al., 2013
Reason #1: Imperfect
Knowledge About Current
State of Atmosphere
Data voids
Interesting weather in between observing locations
Instrument sensitivity and errors
No observations of important quantities
The State of the Problem
• The atmosphere has a total mass of approximately 5.148 x
1021 g (Trenberth and Smith, 2005)
• The mass of 1 mol of dry air is 29.87 g
• This leads to 1.78 x 1020 mol of air
• Each mol of air has Avogadro’s number of particles in it6.022 x 1023 atoms
• So, there are 1.07 x 1044 particles of air in the atmosphere > And we can’t exactly measure these directly anyways!
Reason #2: Imperfect
Computer Model
Need faster, more powerful computers for more grid
points
Reason #2: Imperfect
Computer Model
Need faster, more powerful computers for more grid
points
Need better understanding of physical processes
Many processes in models are parameterized.
Simplifies physical processes that we don’t fully
understand or don’t have measurements of
Calculates processes occurring on scales smaller
than grid boxes
No Man’s horizontal
Land grid spacing
PBL = planetary boundary layer
LES = large-eddy simulation
(Joe Klemp)
What cannot be resolved must be parameterized!
No Man’s Land
(Joe Klemp)
Philosophy of Forecasting
Peter Clark, UK Met Office
What is the phenomenon producing the hazard?
Can our models directly represent it?
Yes: Detection and presentation
No: Diagnostic parameterization
Why parameterize?
• Unresolved physical processes need to be
included in NWP models
• Parameterized quantities often reflect the
sensible weather (clouds, precipitation, surface
temperature, near-surface wind)
• Parameterizations “distill only the essential
aspects of the physical processes they
represent” (Stensrud 2007, p. 9)
Challenge
• Find relationships between subgrid-scale processes
and model-predicted (grid-scale) variables to “close”
the parameterizations.
• This process is reductionist — we assume we can
explain the whole as a sum of its parts.
– We separate boundary layer and land surface processes,
separate boundary layer and shallow cumulus processes,
and represent each one separately.
– We apply the resulting schemes individually, and the
summed outcome is assumed correct!
Stensrud
Subgrid-scale physical processes act within the
vertical column of each grid cell.
Stensrud
The vertical region affected by each scheme varies.
What measurements are needed for NWP?
What observations are routinely available
for the following?
– soils
– vegetation
– boundary layer
– oceans and lakes
– convection
Stensrud
Measurements
– soils
– Temperature, moisture, soil type
– vegetation
– Leaf area (radiation), transpiration rate, moisture content, plant type
– boundary layer
– Turbulent kinetic energy; temperature, moisture, and wind profiles;
aerosol content
– oceans and lakes
– Surface temperature, roughness, surface wind speed, salinity
– convection
– Hydrometeors (type, size, phase, crystal habit), winds
Observations
– soils
– Few and far between.
– vegetation
– AVHRR, MODIS available but not often used
– boundary layer
– Boundary layer depth estimated by soundings, seen in 915-MHz
profilers. No turbulence measurements.
– oceans and lakes
– Satellite data, buoys, ships (little below surface)
– convection
– Radar, satellite
Stensrud
Many of the quantities needed to initialize or verify
parameterization schemes are not routinely
available.
–
–
–
–
–
–
–
–
–
Soil moisture, soil temperature, surface fluxes
Turbulent kinetic energy
Boundary layer depth
Aerosol concentrations, cloud water path
Radiation amounts for all components
Ground temperature, water temperature
Vegetation coverage and biomass
Microphysical particle size distributions
Convective heating profiles
Stensrud
Field observations are usually required to develop
new parameterization schemes.
• Special observations are used to tease
out relationships needed to close the
schemes (relate them to model
variables).
• Schemes are then applied to every grid
point in the model domain.
Stensrud
Parameterizations in NWP models
• Land surface–
atmosphere
– Surface energy budget,
and sensible and latent
radiation
fluxes
• Soil–vegetation–atmosphere
• Water–atmosphere
• Planetary boundary layer
and turbulence
• Convection
• Cloud microphysics
• Clear-sky radiation
• Cloud-cover and cloudysky radiation
• Orographic drag
convection and
microphysics
cloud cover
boundary layer
soil-vegetation
Stensrud
Reason #3: Chaos
“Butterfly effect”
Sensitive dependence to initial conditions
Prof. Ed Lorenz
time 
(Verlaan and Heemink 2001)
Prof. Ed Lorenz
time 
(Verlaan and Heemink 2001)
Prof. Ed Lorenz
time 
bifurcation
(Verlaan and Heemink 2001)
Prof. Ed Lorenz
Small differences in initial conditions
will lead to large differences later.
time 
bifurcation
(Verlaan and Heemink 2001)
Prof. Ed Lorenz
Small differences in initial conditions
will lead to large differences later.
end
start
bifurcation
Prof. Ed Lorenz
Even with a perfect model starting
with perfect initial conditions…
Prof. Ed Lorenz
Even with a perfect model starting
with perfect initial conditions…
weather forecasting is
limited to two weeks.
Given all these limitations,
we have no right to do so well in
forecasting the weather!
What other discipline forecasts the future
with as much success as meteorology?
Improved Computer
Forecasts =
Increasing amount and better
use of data
+
Improving numerical models
+
Ensemble prediction systems
+
Increasing resolution
Improved Computer
Forecasts =
Increasing amount and better
use of data
+
Improving numerical models
+
Ensemble prediction systems
+
Increasing resolution
Members of an
ensemble use slightly
different initial
conditions.
Observation
Ensemble
mean
Spread is a measure
of uncertainty in the
forecast.
Ensemble
mean
Observation
Spread is a measure
of uncertainty in the
forecast.
2°C 2°C
Day1 2
3
4
Spread is a measure
of uncertainty in the
forecast.
2°C 2 9
Day1 2
3
4
5
11°C
6
Spread is a measure
of uncertainty in the
forecast.
2°C 2 9
Day1 2
3
4
5
11°C
6
21°C
13
European Centre for Medium-Range Weather Forecasts
84-h Forecast of Sea-Level Pressure
51-member ensemble: “Postage stamp plot”
European Centre for Medium-Range Weather Forecasts
84-h Forecast of Sea-Level Pressure
Similar forecasts
Same cluster
51-member ensemble: “Postage stamp plot”
European Centre for Medium-Range Weather Forecasts
84-h Forecast of Sea-Level Pressure
Different forecasts
Different clusters
51-member ensemble: “Postage stamp plot”
US National Weather
Service Ensemble
“Spaghetti plot”
of two contours of
path of jet stream
US National Weather
Service Ensemble
Less certainty
“Spaghetti plot”
of two contours of
path of jet stream
More certainty
Met Office Ensemble
Olympic Showcase
>95%
>95%
Probability (%) that rain (>0.2
mm/h) will fall sometime within 18
h
<5%
US National Weather Service National Hurricane Center
US National Weather Service National Hurricane Center
The Goal of US National Weather Service: “Warn on Forecast”
(Hirschberg et al. 2011)
US Navy
Fleet Numerical
Meteorology and
Oceanography Center
HIGH RISK
TO PIRATES
84-h forecast of risk to
pirates in small boats due to
adverse winds and seas
LOW
RISK
(Hirschberg et al. 2011)
Themes for Today
• In principle, weather forecasting is easy.
– Initial state of atmosphere (observations)
– Laws of atmosphere (physics)
Themes for Today
• In principle, weather forecasting is easy.
– Initial state of atmosphere (observations)
– Laws of atmosphere (physics)
• In practice, it is more difficult.
– Errors in initial state
– Approximations
– Chaos
Themes for Today
• In principle, weather forecasting is easy.
– Initial state of atmosphere (observations)
– Laws of atmosphere (physics)
• In practice, it is more difficult.
– Errors in initial state
– Approximations
– Chaos
• Given the extent of the unknown initial state and
the approximations used to solve the equations,
that we succeed so well in weather forecasting is
an amazing human accomplishment.