Mesoscale Probabilistic Prediction over the Northwest: An Overview Cliff Mass University of Washington

Mesoscale Probabilistic Prediction By the late 1990’s, we had a good idea of the benefits of high resolution. It was clear that initial condition and physics uncertainty was large. We were also sitting on an unusual asset due to our work evaluating major NWP centers: real- time initializations and forecasts from NWP centers around the world. Also, inexpensive UNIX clusters became available.

Resolution ( 45  N ) Objective Abbreviation/Model/Source Type Computational Distributed Analysis avn, Global Forecast System (GFS), SpectralT254 / L641.0  / L14 SSI National Centers for Environmental Prediction~55 km~80 km3D Var cmcg, Global Environmental Multi-scale (GEM), Finite0.9  0.9  /L  / L113D Var Canadian Meteorological Centre Diff ~70 km ~100 km eta, limited-area mesoscale model, Finite32 km / L45 90 km / L37SSI National Centers for Environmental Prediction Diff.3D Var gasp, Global AnalysiS and Prediction model,SpectralT239 / L291.0  / L11 3D Var Australian Bureau of Meteorology~60 km~80 km jma, Global Spectral Model (GSM),SpectralT106 / L  / L13OI Japan Meteorological Agency~135 km~100 km ngps, Navy Operational Global Atmos. Pred. System,SpectralT239 / L301.0  / L14OI Fleet Numerical Meteorological & Oceanographic Cntr. ~60 km~80 km tcwb, Global Forecast System,SpectralT79 / L181.0  / L11 OI Taiwan Central Weather Bureau~180 km~80 km ukmo, Unified Model, Finite5/6  5/9  /L30same / L123D Var United Kingdom Meteorological Office Diff.~60 km “Native” Models/Analyses Available

“Ensemblers” Eric Grimit (r ) and Tony Eckel (l) are besides themselves over the acquisition of the new 20 processor athelon cluster

UWME –Core : 8 members, 00 and 12Z Each uses different synoptic scale initial and boundary conditions All use same physics –Physics : 8 members, 00Z only Each uses different synoptic scale initial and boundary conditions Each uses different physics Each uses different SST perturbations Each uses different land surface characteristic perturbations –Centroid, 00 and 12Z Average of 8 core members used for initial and boundary conditions

Ensemble-Based Probabilistic Products

The MURI Project In 2000, Statistic Professor Adrian Raftery came to me with a wild idea: submit a proposal to bring together a strong interdisciplinary team to deal with mesoscale probabilistic prediction. Include atmospheric sciences, psychologists, statisticians, web display and human factors experts.

The Muri I didn’t think it had a chance. I was wrong. It was funded and very successful.

The MURI Over five years substantial progress was made: –Successful development of Bayesian Model Averaging (BMA) postprocessing for temperature and precipitation –Development of both global and local BMA –Development of grid-based bias correction –Completion of several studies on how people use probabilistic information –Development of new probabilistic icons.

Raw 12-h ForecastBias-Corrected Forecast

*UW Basic Ensemble with bias correction UW Basic Ensemble, no bias correction *UW Enhanced Ensemble with bias cor. UW Enhanced Ensemble without bias cor Skill for Probability of T 2 < 0°C BSS: Brier Skill Score

Calibration Example-Max 2-m Tempeature (all stations in 12 km domain)

The Thanksgiving Forecast h forecast (valid Thu 10AM) 13: avn* 11: ngps* 12: cmcg* 10: tcwb* 9: ukmo* 8: eta* Verification 1: cent 7: avn 5: ngps 6: cmcg 4: tcwb 3: ukmo 2: eta - Reveals high uncertainty in storm track and intensity - Indicates low probability of Puget Sound wind event SLP and winds

Ensemble-Based Probabilistic Products

Probability Density Function at one point

But you can have too much of a good thing… Providing forecast uncertainty information is good….

MURI Improvements and extensions of UWME ensembles to multi-physics Development of BMA and probcast web sites for communication of probabilistic information. Extensive verification and publication of a large collection of papers. And plenty more…

Before Probcast: The BMA Site

PROBCAST

ENSEMBLES AHEAD

The JEFS Phase Joint AF and Navy project (at least it was supposed to be this way). UW and NCAR main contractors. Provided support to continue development of basic parameters. Joint project with NCAR to build a complete mesoscale forecasting system for the Air Force. For the first few years was centered on North Korea, then SW Asia, and now the U.S.

JEFS Highlights Under JEFS the post-processed BMA fields has been extended to wind speed and direction. Local BMA for precipitation. Development of EMOS, a regression-based approach that produces results nearly as good as BMA. Next steps: derived parameters (e.g., ceiling, visibility)

NSF Project Currently supporting extensive series of human-subjects studies to determine how people interpret uncertainty information. Further work on icons Further work on probcast.

Ensemble Kalman Filter Project Much more this afternoon. 80-member synoptic ensemble (36 km-12 km or 36 km) Uses WRF model Six-hour assimilation steps. Experimenting with 12 and 4 km to determine value for mesoscale data assimilation-AOR in 3D.

Big Picture The U.S. is not where it should be regarding probabilistic prediction on the mesoscale. Current NCEP SREF is inadequate and uncalibrated. Substantial challenges in data poor areas for calibration and for fields like visibility that the models don’t simulate at all or simulate poorly. A nationally organized effort to push rapidly to 4- D probabilistic capabilities is required.

Opinion Creating sharp, reliable PDFs is only half the battle. The hardest part is the human side, making the output accessible, useful, and compelling. We NEED the social scientists. Probabilistic forecast information has the potential for great societal economic benefit.

The END

Brief History Local high-resolution mesoscale NWP in the Northwest began in the mid-1990s after a period of experimentation showed the substantial potential of small grid spacing (12 to 4 km) over terrain. At that time NCEP was running 32-48km grid spacing and the Eta model clearly had difficulties in terrain.

The Northwest Environmental Prediction System Beginning in 1995, a team at the University of Washington, with the help of colleagues at Washington State University and others have built the most extensive regional weather/environmental prediction system in the U.S. It represents a different model of how weather and environmental prediction can be accomplished.

Pacific Northwest Regional Prediction: Major Components Real-time, operational mesoscale environmental prediction –MM5/WRF atmospheric model –DHSVM distributed hydrological model –Calgrid Air Quality Model –A variety of application models (e.g., road surface) Real-time collection and quality control of regional observations.

Why Probcast? We are rapidly gaining the ability to produce useful probabilistic guidance-- forecasts that are reasonably reliable and sharp. Believe it or not, this is the easy part of the problem. What we have not done is to design interfaces that allow users to make effective user of probabilistic output … or even convince users that they should “go probabilistic”. The recent NRC Report on Probabilistic Prediction highlights this issue.

UW Uncertainty MURI The DOD-sponsored UW Uncertainty MURI was designed to consider both sides of the problem: –Generation of probabilistic information-- ensembles and post-processing –Display and human interface issues. Includes UW Atmospheric Sciences, Statistics, Psychology, and Applied Physics Lab

UW Probabilistic Prediction UW Ensemble System is based on using varying initialization and boundary conditions from differing operational analyses. Also includes varying model physics and surface properties (e.g., SST). Have developed sophisticated post-processing: grid-based bias correction and Bayesian Model Averaging (BMA)--both global and local

Before Probcast: The BMA Site

UW MURI Considerable work by Susan Joslyn and others in psychology and APL to examine how forecasters and others process forecast information and particularly probabilistic information. One example has been their study of the interpretation of weather forecast icons.

The Winner

WRF Domains: km

Putting it All Together During the last year, the MURI project has attempted to create a user interface that accesses the voluminous and detailed probabilistic information produced by our system in a way that is highly accessible to lay and even professional users. The result: PROBCAST Available operationally at

PROBCAST

Probcast Features Local BMA for temperature Currently non-local BMA for precipitation (working on it) Attempt to use probabilistic information in a way that is accessible to deterministic- oriented individuals. Can click anywhere on map to get probabilistic information for specific locations.

PROBCAST

Just the beginning Will expand probcast to more parameters Will improve interface from public feedback Will be testing more icon ideas this year. Unfortunately, MURI ends in 8 months and no replacement.

61 Bayesian model averaging whereis the deterministic forecast from member k, is the weight associated with member k, and is the estimated density function for y given member k

The END

63 Ensemble Forecasting 48-hour forecasts for maximum wind speeds on 7 August 2003

Update on the UW EnKF Mesoscale Analysis and Prediction System Brian Ancell, Clifford F. Mass, and Gregory J. Hakim The University of Washington

Outline 1)Brief review of motivation for a high- resolution EnKF 2)Summary of 36km/12km EnKF results 3)Upgrades to EnKF system 4)Computational costs for more frequent data assimilation cycles at higher resolution

Motivation for High-resolution EnKF Key benefit is using flow-dependent, small- scale structure during assimilation

36-km vs. 12-km EnKF SLP, 925-mb temperature, surface winds 36- km 12- km

36-km vs. 12-km EnKF SLP, 925-mb temperature, surface winds 36- km 12- km

36-km vs. 12-km EnKF 36- km 12- km SLP, 925-mb temperature, surface winds

Motivation for High-resolution EnKF Key benefit is using flow-dependent, small- scale structure during assimilation Analysis and forecast uncertainty is a major advantage

EnKF Configuration D1 (36km) D2 (12km) D3 (4km)

EnKF Configuration WRF model 38 vertical levels 80 ensemble members 6-hour update cycle Observations: Surface temperature, wind, altimeter ACARS aircraft winds, temperature Cloud-track winds Radiosonde wind, temperature, relative humidity Half of surface obs used for assimilation, other half for verification

EnKF 12km Surface Observations

EnKF 36km vs. 12km Improvement of 12-km EnKF Analys is Foreca st 10 % 13 % WindTemperatur e

Reduction of Observation Variance in 12km EnKF Improvement using reduced observation uncertainty Analys is 5%10 % WindTemperatur e

Further Improvement After Variance Inflation, Bias Removal Improvement using inflation and bias removal Analys is 9%3% WindTemperatur e

EnKF 12-km vs. GFS, NAM, RUC RMS analysis errors GF S 2.38 m/s 2.28 K NA M RU C EnKF 12-km WindTemperatur e 2.30 m/s 2.13 m/s 1.85 m/s 2.54 K 2.35 K 1.67 K

EnKF System Upgrades 4km domain

EnKF System Upgrades 36k m 12k m 4km SLP, 925-mb temperature, surface winds

EnKF System Upgrades 4km domain WRF V2 to WRF V3

EnKF System Upgrades 4km domain WRF V2 to WRF V3 Physics matched with real time GFS-WRF (changes bias and surface variance characteristics)

EnKF System Upgrades 4km domain WRF V2 to WRF V3 Physics matched with real time GFS-WRF (changes bias and surface variance characteristics) Additional altimeter observations Surface observations 12km: ~10,000 ~5,000 (50%) 4km: ~5,000 ~4,000 (80%) Terrain check

EnKF System Timing 6-hr cycle, 80 processors 36km: assimilation: 12 minutes model integration: 20 minutes overhead: 5 minutes 4km assimilation: 15 minutes model integration: 240 minutes nestdown: 10 minutes overhead: 10 minutes Total: 5 hours

EnKF System Timing 3-hr cycle, 80 processors 36km: assimilation: 12 minutes model integration: 10 minutes overhead: 5 minutes 4km assimilation: 15 minutes model integration: 120 minutes nestdown: 10 minutes overhead: 10 minutes Total: 3 hours

EnKF System Timing 2-hr cycle, 80 processors 36km: assimilation: 12 minutes model integration: 7 minutes overhead: 5 minutes 4km assimilation: 15 minutes model integration: 80 minutes nestdown: 10 minutes overhead: 10 minutes Total: 2.3 hours

EnKF System Timing 2-hr cycle, 160 processors 36km: assimilation: 12 minutes model integration: 4 minutes overhead: 5 minutes 4km assimilation: 15 minutes model integration: 50 minutes nestdown: 5 minutes overhead: 10 minutes Total: 1.7 hours