The Centre for Australian Weather and Climate Research A partnership between CSIRO and the Bureau of Meteorology Alan Seed Centre for Australian Weather

The Centre for Australian Weather and Climate ResearchA partnership between CSIRO and the Bureau of Meteorology

Alan Seed

Centre for Australian Weather and Climate Research

Short Term Ensemble Prediction System: STEPS

The Centre for Australian Weather and Climate Research A partnership between CSIRO and the Bureau of Meteorology

Outline

• Statistical structure of rainfall

• Modelling the errors in a nowcast• Temporal development • Radar reflectivity to rain rate conversion• Tracking

• Nowcast ensembles• Radar only nowcasts• Radar + NWP blending

• Products• Ensembles for end users• Expected rainfall – ensemble mean• Probability of exceeding various thresholds• Meteograms

• Products

• Developments

• Conclusions


15-min rainfall over the UK 1000 km (15 min)


Adelaide Radar – 250 km (10 min)


Auckland 10 km (2 min)

15 km x 7.5 km box, 100 m resolution


Variability as a function of scaleM

od

elli

ng

1000 km domain, eastern half of the HRRR region

2d power spectrum

0

10

20

30

40

50

60

70

80

-30 -20 -10 0

10log(1/km)

10

log

(po

we

r)

mit vil

HRRR


Nowcast skill as a function of scale and lead time

Widespread rain in Sydney

Skill vs Lead Time

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1 2 3 4 5 6 7

Lead Time (hours)

Co

rrel

atio

n

500 km

250 km

100 km

50 km

20 km

10 km


Outline

• Statistical structure of rainfall• Modelling the errors in a nowcast

• Temporal development • Radar reflectivity to rain rate conversion• Tracking



• Developments• Conclusions


Conceptual model for rainfall

• Rainfall usually has areas of higher intensity rainfall inside areas of lower intensity rainfall, and we get clusters of storms and not just a random pattern of storms- variability over a wide range of scales

• The lifetime of a storm increases with the size of the storm as a power law

• The simplest model is a multiplicative cascade model (used to model turbulence) for the spatial scaling and a hierarchy of AR(1) models for the Lagrangian temporal evolution so as to reproduce the dynamic scaling of the field

Tem

po

ral

dev

elo

pm

ent

of

rain

fall


Multiplicative Cascade Model for Turbulence

Lovejoy et al., 1987

J. Geophys. Res.

Each cascade level evolves in time Rate of development decreases with increasing scaleHierarchy of AR(1) models used for temporal development

Tem

po

ral

dev

elo

pm

ent

of

rain

fall


Spectral decomposition of a rainfall field

128-256-512 km

64-128-256 km

32-64-128 km

16-32-64 km 8-16-32 km 2-4-8 km4-8-16 km

Tem

po

ral

dev

elo

pm

ent

of

rain

fall


Radar Z-R error is coherent over scales that are significant for urban hydrology

Radar Gauge

Rad

ar m

easu

rem

ent

erro

rZ

- R

Err

or


Space & Time correlations of radar z-r error Villarini et al, WRR 45, W01404 2009

Z -

R E

rro

r


Errors due to observing above ground level

Correlation below WBFL

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

0.00 0.50 1.00 1.50 2.00 2.50 3.00 3.50 4.00

Height (km)

Co

rrel

atio

n Bris rain

Melb rain

Kur rain

Model

Correlation as a function of height separation for pairs of radar observations where one observation is at the base scan and the other is below the wet bulb freezing level.

Sam

pli

ng

Err

or


Model of radar ZR and sampling errorsObserved and ensembles for 10-min rainfall

Rad

ar m

easu

rem

ent

erro

r

Ensemble 1 Ensemble 2

Observation

Radar observation error model includes Z-R and sampling errors due to observing at a height above the ground M

od

elli

ng

QP

E E

rro

r


Verification of radar error model

-30

-20

-10

0

10

20

30

-35 -30 -25 -20 -15 -10 -5 0

10log(frequency)

po

wer Model

Radar

Reliability of probabilities Power specta of observed and perturbed fields

Mo

del

lin

g Q

PE

Err

or


Other radar observation errors are tricky, depending on the situation and the QC algorithms used

Rad

ar m

easu

rem

ent

erro

r

Beam blocking

Clutter

Daily rainfall accumulation for Melbourne

Mo

del

lin

g Q

PE

Err

or


Modelling the tracking error

• We do not have a complete description of tracking error• Generating fields of U,V error components that are correlated with

each other and in space and time is VERY tricky- at least I do not know how to do it

• Not the most important source of error in the first 6 hours so we can keep it simple

• Multiply the radar U,V components by a random number that has a mean of 1 and some (small) variance

Mo

del

lin

g T

rack

ing

Err

or


Outline


• Temporal development • Radar reflectivity to rain rate conversion• Tracking





Short Term Ensemble Prediction System- radar only

1. Estimate the advection field using rainfall fields

2. Estimate the AR(1) and cascade parameters using the current observed field

3. For each ensemble membera) Perturb the radar analysis with the observation error

model

b) Perturb the advection field

c) Generate a conditional stochastic field for the next 90 minutesM

od

elli

ng

ST

EP

S-n

ow

cast


Beijing Olympics: 1 hr forecast & observationS

TE

PS

-no

wca

st


Beijing 2008 inter-comparison

• Compared STEPS against 5 other international systems during the Beijing Games

ST

EP

S-n

ow

cast


Short Term Ensemble Prediction System- NWP blend

• Decompose NWP into a cascade • Decompose the rainfall field into a cascade• Use radar field to estimate stochastic model parameters• Calculate the skill of the NWP at each level in the

cascade using the correlation between NWP and radar• Blend each level in the radar & NWP cascades using

weights that are a function of the forecast error at that scale and lead time

• For each forecast • Add noise component to the deterministic blend, the

weight of the noise is calculated using the skill of the blended forecast

• Combine the cascade levels to form a forecast

ST

EP

S-N

WP


Blending with NWP – calculating the weights

bb2b-)var(

variancehas noise remaining

truth with theforecasts of covariance theis

forecasts theof cavariance theis

weightsof vector theis

and blended, be toforecasts of vector theis

forecast blended theis y

)(ˆ

XXT

XYT2

XY

1

YY

XX

XYXX

TxY

Z

b

b

X

bXy

ST

EP

S-N

WP


Nowcast explained variance as a function of scale and lead time

Extrapolation forecast

0.00

0.20

0.40

0.60

0.80

1.00

0 10 20 30 40 50 60 70 80 90 100 110 120

Lead time (min)

Exp

lain

ed v

aria

nce

736 Extrapolation 275 Extrapolation

100 Extrapolation 40 Extrapolation

ST

EP

S-N

WP


NWP explained variance as a function of scale

NWP forecast

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

0 100 200 300 400 500 600 700 800

Scale (km)

Exp

lain

ed v

aria

nce

ST

EP

S-N

WP


Weights for nowcast & NWPB

len

din

g

500 km scale

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 100 200 300 400

Lag (minutes)

Wei

gh

t Extrapolation

NWP

Noise

125 km scale

0.00

0.20

0.40

0.60

0.80

1.00

1.20

0 100 200 300 400

Lag (minutes)

Wei

gh

t Extrapolation

NWP

Noise

ST

EP

S-N

WP


Outline


• Radar reflectivity to rain rate conversion• Tracking• Temporal development





Status

• QPE and STEPS-nowcast running on LINUX workstations in operational mode

• STEPS-NWP (radar + NWP blend) running on a super computer • 16 radars with QPE, 15 QPF domains• Generating 1000 products (100 Mbytes) per hour• Up to 100 clients inside the Bureau being served with products on a

busy day• QPE live to the public for capital city radars• Planning to go live to the public with QPF in May 2011

Pro

du

cts


Rainfall Estimates

• Melbourne, Sydney, Brisbane, Adelaide • 30, 60, 120 min, since 9 AM, daily accumulations blended

with rain gauges and updated every 30 min • 10 min accumulations radar only with real-time gauge

adjustments and updated every 6 or 10 minutes

Pro

du

cts


Rainfall Forecasts: 0 – 90 minutes

• 4 major cities, 1 km & 6 min resolution, 250 km domain• 3 Regional forecasts, 2 km & 10 min resolution, 500 km domain• 30 member ensemble updated every 6,10 minutes• 30, 60, 90 min accumulations of ensemble mean (expected rain)• Probability that rain accumulation will exceed 1,2,5,10,20,50 mm in

next 60 minutes

60 min accumulation Probability of rain > 50 mm Forecast time series at a pointwith uncertainty shown

Pro

du

cts


Rainfall forecasts: 1 – 6 hours

• Melbourne, Sydney, Brisbane – 500 km domain, 2 km & 10 min resolution

• 30 member ensemble updated every hour• 10-min forecasts of rainfall intensity out to 6 hours• Probability products for hourly accumulations for next 6 hours

Probability of rain > 1 mm for 2 & 3 hour lead times, Melbourne

Rainfall intensity forecast, 150 min lead time, Brisbane

Pro

du

cts


Example from east Victoria – NWP


Example from east Victoria- STEPS

Ensemble member 1


Met Service Canada: Point Mode

Paul Joe, 2010


Met Service Canada: Point-Time Mode

PDF of rainrates at a point for all time lagged nowcastsPaul Joe, 2010


Met Service Canada: POP for a validtime and rainrate threshold

Rainrate threshold is 1 mm/h; number of hits exceeding threshold / number of samples (60)

Paul Joe, 2010


Probability

Probability of 60 min accum > 5 mm


Outline


• Radar reflectivity to rain rate conversion• Tracking• Temporal development





Qualitative severe weather warning

• THESPA • Calculates the probability of a TITAN cell passing over a point in the next 60

minutes based on the current velocity and cell size and a climatological TITAN tracking error

• Being developed for aviation applications and use in TIFS

• TIFS • Operational in most Regional Forecast Centres

• Automatic version for aviation is operational

• Revising the software architecture

• Graphical and automated text editing feature development

25 km

Dev

elo

pm

ents


Development: Closing the gap between NWP & nowcasts

• Strategic Radar Enhancement Project • $48 M project over 7 years, 8 people for three years in CAWCR

• 4 new radars

• Radar data assimilation in ACCESS• Roll out of a new radar data quality control system for ~50 radars• Characterise radar errors for use in data assimilation (and QPE)• Assimilate radar data (LH nudging, Doppler radial winds, reflectivity) into high res

(~2 km) NWP meso-scale models over capital cities

• Seamless rainfall prediction• Integrate rainfall forecasts from 0 – 10 days lead time into a seamless

forecast

• Use STEPS to blend the forecasts from the various models

• Develop a portal for convenient access to the rainfall forecasts


Seamless rainfall forecasts

• Rainfall portal

• Data source transparent to user

• Aggregation

• Disaggregation

• New science

• STEPS downscaling

• Blending strategy

• Verification (esp. transition periods)

3-10 day forecasts

40-100 km

2 x daily

1-2 day forecasts

ACCESS-AACCESS-CAGREPS-RECMWF EPSPME/GOCFGFE

5-25 km

2-4 x daily

1-24 hour forecasts

ACCESS-AACCESS-CGFE

2-10 km

4-8 x daily

1-6 hour forecasts

ACCESS-CSTEPS-NWP

1-2 km

hourly

10-90 min nowcasts STEPS-nowcast every 10

min1 km

Down-scaled and blended using STEPS

ACCESS-GAGREPS-GECMWF EPSPME/GOCFGFETIGGE?

C

ross

-cut

ting

prog

ram

s: E

SM

, C

WD

, O

EB

, N

MO

C


Issues

• Limited capacity in the Regional Forecast Centres Head Office support branches to deploy and learn new nowcasting systems, busy with the Next Generation Forecast and Warning System – slows the adoption of new algorithms

• Focus has been on improving the service adoption of existing nowcasting science through

• Delivering the products through a range of platforms – 3drapic, Google maps, web pages

• Using formats that are carefully designed and that conform to formal geo-spatial standards (eg CF compliant netCDF)

• Serving the data on a range of platforms (ftp, SOAP, directories)

• Formalising the use of QPE&F products in the forecast process

• Training

Dev

elo

pm

ents


Conclusions

• Nowcasting rainfall is an uncertain business • Have incomplete description of the error structure of QPE and QPF• Have enough of a description to make useful stochastic ensemble

models• There is still a lot of work to do to make the stochastic models include

more meteorological knowledge• There is even more work to do to help the end-users make full use of

the ensemble members in their decision support systems

Documents

The Centre for Australian Weather and Climate Research A partnership between CSIRO and the Bureau of Meteorology Alan Seed Centre for Australian Weather