Ultra Rapid Data Assimilation for Real Time Weather Walter Acevedo, - - PowerPoint PPT Presentation

Ultra Rapid Data Assimilation for Real Time Weather

Walter Acevedo, Zoi Paschalidi, Christian Welzbacher & Roland Potthast January 2019 ISDA 2019, Kobe

Motivation

Autonomous driving

depends strongly on weather conditions!

page 2

• Motivation
• New observation sources
• Ultra-Rapid DA (URDA)

Algorithm

• Experiments with KENDA

system

• Conclusions and

perspective

Source: AUDI AG

Outline

• Motivation
• New observation sources
• Ultra-Rapid DA (URDA)

Algorithm

• Experiments with KENDA

system

• Conclusions and

perspective

Source: AUDI AG

Outline

New observation sources

Car observations

Source: AUDI AG

page 4

New observation sources

page 5

Car observations

Source: AUDI AG

New observation sources

Measured Variables § Road temperature § Road condition § Air temperature § Dew point § Temperature at 30 cm depth § Precipitation type § Precipitation intensity § Wind direction § Wind velocity § Visibility

Road Weather Stations

Source: Bavarian Administration

page 6

New observation sources

Road Weather Stations

Source: Bavarian Administration

Coverage for Bavaria on January 2017 (285 stations)

page 7

• Motivation
• New observation sources
• Ultra-Rapid DA (URDA)

Algorithm

• Experiments with KENDA

system

• Conclusions and

perspective

Source: AUDI AG

Outline

Analysis Cycle:

𝑢" 𝑢# 𝑢"$%

• Forecast step
• Assimilation step

Normal sequential DA

Model Dynamics Forecast Ensemble-Transformation Analysis New Analysis New Forecast

page 8

Ultra-Rapid DA (URDA)

𝑢" 𝑢# 𝑢"$%

„Preemptive Forecasting“:

page 9

• 2007:
• 2015:
• 2018:

Properties:

• Equivalent to Normal sequential DA for linear Model and observation
• perator [Potthast & Welzbacher 2018]:
• Applicable to several assimilation steps from different time intervals:
• No model reinitialization necessary.
• No need to update the whole model state. A reduced set of selected

variables and gridpoints can be updated:

Ultra-Rapid DA (URDA)

page 10

• Motivation
• New observation sources
• Ultra-Rapid DA (URDA)

Algorithm

• Experiments with KENDA

system

• Conclusions and

perspective

Source: AUDI AG

Outline

Experiments with KENDA system

page 11

COSMO-D2:

• Limited-area short-range convection-

permitting numerical model weather prediction model

• Dx @ 2.2 km / 65 vertical layers
• Explicit deep convection

Local Ensemble Transform Kalman Filter:

• LETKF implementation following Hunt et

al., 2007

• Rapid Update cycle (RUC) of 1 hour
• 40 members + 1 deterministic run
• adaptive horizontal localisation
• adaptive multiplicative inflation + RTPP
• additive covariance inflation

KENDA (Km-scale ENsemble DA) system:

„Model state“

RUC run URDA run Free run

01:00 02:00

Observation

W W W W W W W

15 min 10 min 5 min

00:00

RUC-run

Forecast time :

URDA-KENDA Experiment

* RUC = rapid update cycle – 1h assimilation cycle page 12

• Three conventional observational sources:

1. Synoptic stations 2. Radiosondes 3. Aircraft observations

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

URDA-KENDA Experiment

Pa Pa

Unbalance introduced Stable behaviour

Pressure Differences (ground layer) regarding the free forecast

URDA-KENDA Experiment

RUC minus Free run URDA minus Free run

page 13

Pa Pa

Spinup waves spread out

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 13

Stable behaviour Spinup waves spread out

URDA-KENDA Experiment

Pa Pa

Pressure Differences (ground layer) regarding the free forecast

Stable behaviour

RUC minus Free run URDA minus Free run

page 13

Spinup waves spread out

URDA-KENDA Experiment

C° C°

Spinup-Effect

Temperature Differences (ground layer) regarding the free forecast

RUC minus Free run URDA minus Free run

page 14

Stable behaviour

URDA-KENDA Experiment

„Model state“

RUC-run URDA-run Free run

01:00 02:00

Observation

W W W W W W W

15 min 10 min 5 min

00:00

RUC-run

Forecast time :

• Model re-initialization introduces imbalances
• URDA can beat RUC for short forecast times

Spinup-Periode Unbalance model state

* RUC = rapid update cycle – 1h assimilation cycle page 15

URDA-KENDA Experiment

Pressure improved

Assimilating Pressure observations ONLY

page 16

Root Mean Square Error (RMSE) vs Leadtime

URDA-KENDA Experiment

Pressure improved

Assimilating Pressure, Temperature and U-V Wind observations

page 17

Root Mean Square Error (RMSE) vs Leadtime

V-wind improved U-wind improved

URDA-KENDA Experiment

• Motivation
• New observation sources
• Ultra-Rapid DA (URDA)

Algorithm

• Experiments with KENDA

system

• Conclusions and

perspective

Source: AUDI AG

Outline

Conclusions and Perspective

1. URDA shows great potential for short leadtimes:

• Reduced computational cost
• No spinup effect

2. Observation operator for car observations under development

• Quality control for a moving weather station
• Modelling of dependency between meteorological state and

auto-microclimate

• Auto-dependent bias correction
• Time- and spatial aggregation
• Data anonymization

page 18

References

1.

• B. J. Etherton. Preemptive forecasts using an ensemble

kalman filter. Monthly Weather Review,135(10):3484–3495, 2007. 2.

• L. E. Madaus and G. J. Hakim. Rapid, short-term ensemble

forecast adjustment through offline data assimilation. Quarterly Journal of the Royal Meteorological Society, 141(692):2630–2642,2015. 3.

• R. Potthast and C.A. Welzbacher. Ultra Rapid Data

Assimilation Based on Ensemble Filters. Front. Appl. Math.

• Stat. 4:45, 2018.

4. Hunt BR, Kostelich EJ, and Szunyogh I. 2007. Efficient data assimilation for spa-tiotemporal chaos: A local ensemble transform Kalman Filter.Physica D,230: 112-126.

page 19

Vielen Dank! Thank you! Gracias! Eυχαριστώ!

どうもありがとう