SUB-SEASONAL FORECASTING USING LARGE ENSEMBLES OF DATA-DRIVEN GLOBAL WEATHER PREDICTION MODELS
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
SUB-SEASONAL FORECASTING USING LARGE ENSEMBLES OF
DATA-DRIVEN GLOBAL WEATHER PREDICTION MODELS
JONATHAN WEYN1*, DALE DURRAN1, RICH CARUANA2
1 University of Washington, Seattle, WA
2 Microsoft Research, Redmond, WA
* Current affiliation: Microsoft
AI for Earth
of Newtonian determinism, is all the more audacious given that, at that nomenon can be predicted 3–4 months ahead. Weather and climate
time, there were few routine observations of the state of the atmosphere, prediction skill are intimately linked, because accurate climate predic-
INTRODUCTION
no computers, and little understanding of whether the weather possesses tion needs a good representation of weather phenomena and their stat-
any significant degree of predictability. But today, more than 100 years istics, as the underlying physical laws apply to all prediction time ranges.
later, this paradigm translates into solving daily a system of nonlinear This Review explains the fundamental scientific basis of numerical
NWP:
differential equations “Q
A at aboutUIET R EVOLUTION
half a billion points ”
per time step between weather prediction (NWP) before highlighting three areas from which
the initial time and weeks to months ahead, and accounting for dynamic, the largest benefit in predictive skill has been obtained in the past—
thermodynamic, radiative and chemical processes working on scales physical process representation, ensemble forecasting and model initi-
from hundreds of metres to thousands of kilometres and from seconds alization. These are also the areas that present the most challenging
to weeks.
• Weather forecasting has gradually increased in science questions in the next decade, but the vision of running
A touchstone of scientific knowledge and understanding is the ability
accuracy, due to:
to predict accurately the outcome of an experiment. In meteorology, this Day 3 NH Day 5 NH Day 7 NH Day 10 NH
translates Vastthe
• into advances
accuracy ofinthe
numerical representation
weather forecast. of
In addition, today’s Day 3 SH Day 5 SH Day 7 SH Day 10 SH
98.5
numerical atmospheric
weather predictions also enable
physics the forecastermethods
and numerical to assess quan-
titatively the degree of confidence users should have in any particular 95.5
forecast. Large
• This network
is a story of data
of profound andcollection
fundamental from satellites
scientific success 90
Forecast skill (%)
built upon the application of the classical laws of physics. Clearly the
• Supercomputing power
success has required technological acumen as well as scientific advances 80
Is it possible to stretch these improvements even
and• vision. 70
further
Accurate by applying
forecasts save lives,modern machine learning
support emergency management and 60
mitigation of impacts and prevent economic losses from high-impact
techniques? 50
weather, and they create substantial financial revenue—for example, in
40
energy, agriculture, transport and recreational sectors. Their substantial
benefits far outweigh the costs of investing in the essential scientific 30
1981 1985 1989 1993 1997 2001 2005 2009 2013
research, super-computing facilities and satellite and other obser-
3 Year
vational programmes that are needed to produce such forecasts . Bauer et al. (2015)
These scientific and technological developments have led to increas- Figure 1 | A measure of forecast skill at three-, five-, seven- and ten-day
ing weather forecast skill over the past 40 years. Importantly, this skill ranges, computed over the extra-tropical northern and southern
can be objectively and quantitatively assessed, as every day we compare hemispheres. Forecast skill is the correlation between the forecasts and the
JONATHAN WEYN, MICROSOFT verifying analysis of the height of the 500-hPa level, expressed as the anomaly
the forecast with what actually occurs. For example, forecast skill in the
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES with respect toFOR
the climatological
ML AND AI IN height.
WEATHER ValuesAND
greater than 60%
CLIMATE indicate
MODELING
range from 3 to 10 days ahead has been increasing by about one day per useful forecasts, while those greater than 80% represent a high degree of 2
INTRODUCTION
EXPLOSION IN APPLICATION OF ML IN METEOROLOGY
• Post-processing NWP models (Rasp and Lerch
2018)
High-resolution truth model
• Identification and prediction of extreme weather Low-resolution model with ML physics
events (Racah et al. 2017, Lagerquist et al. 2019, Low-resolution model
Herman and Schumacher 2018)
• Improving physical parameterizations in NWP
models, and improving their computational
efficiency (Rasp et al. 2018, Brenowitz and
Bretherton 2018, McGibbon and Bretherton 2019)
Rasp et al. (2018)
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
3
INTRODUCTION
WHAT IF MACHINES COULD PREDICT THE EVOLUTION
OF THE ENTIRE ATMOSPHERE?
• We developed our Deep Learning Weather Prediction (DLWP) model
• Use deep convolutional neural networks (CNNs) on a cubed sphere grid
• While this is essentially replacing NWP models with deep learning, there are
important caveats
• only a few variables: not a complete forecast
• subject to using training data dependent on state-of-the-art data assimilation
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
4
INTRODUCTION
Indefinitely iterable predictions for every 6 hours
Observed states Predicted states Predicted states
at two times at two times at two times
x(t – 6)
˜ – 6, t)
x(t Algorithm ˜ + 6, t + 12)
y(t Algorithm ˜ + 18, t + 24)
y(t ...
x(t)
2 2 2 2
y(t + 6) y(t + 12) y(t + 18) y(t + 24)
J1 = – + – J2 = – + –
x(t + 6) x(t + 12) x(t + 18) x(t + 24)
2 2 2 2
J = J1 + J2
Trained to minimize the error over a 24-hour forecast
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
5
DLWP ALGORITHM
CONVOLUTIONAL NEURAL NETWORKS
• CNNs
• are ideally suited for “images” on a grid
• account for local spatial correlations
• identify patterns, edges, shapes
• However, equirectangular (latitude-longitude)
images have heavy distortion near the poles
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
6
DLWP ALGORITHM
CONVOLUTIONS ON THE CUBED SPHERE
a)
310
c)
100 b)
300
50 290
280
T2 (K)
0
270
50 260
250
100
240
0 50 100 150 200 250 300 350
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
7
INTRODUCTION
64
3×3 convolution
2×2 average pooling
2×2 up-sampling
skip connection
1×1
conv
s
Preserves fine-scale information
le
ca
ls
ia
at
128
sp
e
pl
ti
ul
m
at
n
io
at
m
or
nf
256
si
te
ora
rp
co
In
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
8INTRODUCTION
DATA
• ECMWF ERA5 input fields (6) • Prescribed fields (3)
• ~1.4º resolution; 6-hourly time step • TOA incoming solar radiation
• Z500 , Z1000 • land-sea mask
• 300–700 hPa thickness • topography
• 2-m temperature • Data sets
• T850 • Train: 1979–2012; validate: 2013–16
• Total column water vapor • Evaluate: twice weekly in 2017–18
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
9ML MODEL
DLWP produces
realistic, albeit
smoothed,
Colors: 500-hPa geopotential height, dkm
atmospheric
Contours: 1000-hPa geopotential height, every 100 m, dashed=negative
states. https://www.dropbox.com/s/xmi0v81fbcw4t1r/Weyn_ms02.mp4?dl=0
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
10DLWP ENSEMBLE
Can our DLWP model form the basis for a good-performing
large ensemble prediction system?
• DLWP is inferior to NWP models. Why bother?
• The larger the ensemble, the better!
• DLWP has a significant computational
advantage
• How long would it take to run a 1000-
member ensemble of 1-month forecasts at
1.5º resolution?
• 10 minutes
• Single workstation with GPU
• In comparison, a comparable dynamical
model would take about 16 days
Morris MacMatzen / Getty via Vox
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
11DLWP ENSEMBLE
OBTAINING CORRECT ENSEMBLE SPREAD
NWP DLWP
• Perturb initial conditions • Perturb initial conditions
Observation uncertainty • Ensemble 4DVAR • ERA5 ensemble
• SVD • NOT optimal!
• Stochastically perturbed • Random seeds in training
parameterization tendencies • random data sampling
Model uncertainty (SPPT) • random weight
• Stochastic KE backscatter initialization
(SKEB) • “swapping” physics
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
12DLWP ENSEMBLE
DLWP VERSUS THE ECMWF ENSEMBLE
DLWP ECMWF
9 prognostic 3-D variables,
Variables 6 2-D variables
91 vertical levels
Resolution ~160 km ~18 km (36 km after day 15)
Other physics 3 prescribed inputs Many parameterizations
Coupled models None Ocean, wave, sea ice
Initial condition perturbations 10 (ERA5) 50 (SVD/4DVAR)
Model perturbations 32 “stochastic” CNNs stochastic physics perturbations
Ensemble size 320 (+ control) 50 (+ control)
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
130.2
0.5 Average forecast error in 500-hPa geopotential
0.0
0.0 Root-mean-squared error Anomaly correlation coefficient
1.0
5
Worse
Better
a)
e) b)
f)
1000
1000
0.8
4
800
800
0.6
3
Z500 RMSE
T850 RMSE
ACC
ACC
600
600 2–3 days
500
Z850
0.4
T
2
400
400 DLWP control
DLWP grand ensemble
ECMWF S2S control
0.2
1
200
200 ECMWF S2S ensemble
Worse
Better
Persistence
Climatology 0.0
0
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
forecast day 1.0 forecast day
3.5
c) d)
3.0 The DLWP ensemble only lags the state-of-the-art0.8
ECMWF ensemble by 2-3 days’ lead time.
2.5
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES
0.6 FOR ML AND AI IN WEATHER AND CLIMATE MODELING
2.0 14
E0.2
0.5 Average forecast error in 500-hPa geopotential
0.0
0.0 Root-mean-squared error Anomaly correlation coefficient
1.0
5
Worse
Better
a)
e) b)
f)
1000
1000
0.8
4
800
800
0.6
3
Z500 RMSE
T850 RMSE
ACC
ACC
600
600 2–3 days
500
Z850
0.4
T
2
400
400 DLWP control
DLWP grand ensemble
ECMWF S2S control
0.2
1
200
200 ECMWF S2S ensemble
Worse
Better
Persistence
Climatology 0.0
0
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
forecast day 1.0 forecast day
3.5
c) d)
3.0 The DLWP ensemble only lags the state-of-the-art0.8
ECMWF ensemble by 2-3 days’ lead time.
2.5
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES
0.6 FOR ML AND AI IN WEATHER AND CLIMATE MODELING
2.0 14
EDLWP FOR S2S
Can our large ensemble produce skillful sub-seasonal-to-seasonal (S2S) forecasts?
Potential applications of subseasonal-to-seasonal (S2S) predictions 317
(a) WEATHER FORECASTS
predictability comes from initial
atmospheric conditions
S2S PREDICTIONS
predictability comes from initial
atmospheric conditions, monitoring the
land/sea/ice conditions, the stratosphere
excellent and other sources
SEASONAL OUTLOOKS
predictability comes primarily from
FORECAST SKILL good sea-surface temperature conditions;
accuracy is dependent on ENSO state
fair White et al. (2017)
poor
zero
Daily values
1–10 days Weekly averages
10–30 days Monthly or seasonal averages
30–90+ days
FORECAST RANGE
JONATHAN WEYN, MICROSOFT (b) USER NEEDS FOR ML AND AI IN WEATHER AND CLIMATE MODELING
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES
Reliable and actionable information for decision-making 15DLWP FOR S2S
SUB-SEASONAL-TO-SEASONAL FORECASTING
Potential applications of subseasonal-to-seasonal (S2S) predictions
• DLWP cannot be expected to compete with the (a) WEATHER FORECASTS
predictability comes from initial
state-of-the-art atmospheric conditions
S2S PREDICTIONS
• DLWP is lacking predictability comes from initial
atmospheric conditions, monitoring the
• an ocean model excellent
land/sea/ice conditions, the stratosphere
and other sources
SEASONAL OUTLOOKS
• land surface parameters e.g. soil moisture predictability comes primarily from
FORECAST SKILL
good sea-surface temperature conditions;
accuracy is dependent on ENSO state
• sea ice
fair
• We also do not forecast precipitation with the
current model iteration poor
• Cannot represent physics of MJO or ENSO
zero
Daily values
1–10 days Weekly averages
10–30 days Monthly or seasonal averages
30–90+ days
FORECAST RANGE
(b) USER NEEDS
JONATHAN WEYN, MICROSOFT Reliable and actionable information for decision-making
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
16DLWP FOR S2S
PROBABILISTIC SKILL SCORES OF S2S FORECASTS
• Ranked probability skill score (RPSS)
• Three equally-probable terciles relative to a
1981–2010 climatology
• below-, near-, and above-normal
• Forecast probability is the fraction of ensemble
members
• Evaluate squared error of forecast probability
versus observed category
• Normalize relative to random chance prediction
• Perfect score is 1, higher is better, negative
score is no skill relative to random chance
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
17DLWP FOR S2S
PROBABILISTIC SKILL SCORES OF S2S FORECASTS
• Ranked probability skill score (RPSS)
• Three equally-probable terciles relative to a
1981–2010 climatology
• below-, near-, and above-normal
• Forecast probability is the fraction of ensemble
members
Land points only
• Evaluate squared error of forecast probability
versus observed category
• Normalize relative to random chance prediction
• Perfect score is 1, higher is better, negative
score is no skill relative to random chance
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
18While most skill comes from the tropics, the DLWP ensemble
Week 3 has positive skill scores over nearly all land masses. Weeks 5-6
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
19DLWP fails to predict the onset of ENSO patterns in forecasts
Week 3 initialized in boreal autumn. Weeks 5-6
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
20CONCLUSIONS
1.0
a)
1000
0.8
TAKE-AWAYS
800
0.6
Z500 RMSE
Z500 ACC
600
0.4
400 DLWP control
DLWP grand ensemble
ECMWF S2S control
0.2
200 ECMWF S2S ensemble
Persistence
Climatology 0.0
0
3.5 1.0
c)
• Purely data-driven algorithms (CNNs) can produce realistic, indefinitely- 3.0
2.5
0.8
stable global weather forecasts with skill that only lags the state-of-the-art
0.6
2.0
T2 RMSE
T2 ACC
1.5 0.4
ECMWF model by 2-3 days of lead time. 1.0
0.5
0.2
0.0
By leveraging initial-condition perturbations and the internal randomness
0.0
• 5
e)
1.0
of the CNN training process it is possible to produce a high-performing
0.8
4
0.6
3
ensemble of data-driven weather models that requires several orders of
T850 RMSE
T850 ACC
0.4
2
magnitude less computation power than comparable dynamical models. 1
0.2
0.0
0
0 2 4 6 8 10 12 14 0
For weekly-averaged forecasts in the S2S forecast range, a notable skill
forecast day
•
gap for dynamical models, our data-driven ensemble model has useful
skill relative to climatology and persistence. It is not far behind the state-
of-the-art ECMWF ensemble.
JONATHAN WEYN, MICROSOFT
2021 JOINT IS-ENES3/ESIWACE2 VIRTUAL WORKSHOP ON NEW OPPORTUNITIES FOR ML AND AI IN WEATHER AND CLIMATE MODELING
21You can also read
