Ensembling Medium Range Forecast MOS GUIANCE By Richard H. Grumm National Weather Service State College PA 16803 and Robert Hart The Florida State University.

Slides:



Advertisements
Similar presentations
Anticipating Heavy Rainfall: Forecast Aspects By Richard H. Grumm* National Weather Service State College PA and Robert Hart The Pennsylvania State.
Advertisements

ECMWF long range forecast systems
KMA will extend medium Range forecast from 7day to 10 day on Oct A post processing technique, Ensemble Model Output Statistics (EMOS), was developed.
Section #1 October 5 th Research & Variables 2.Frequency Distributions 3.Graphs 4.Percentiles 5.Central Tendency 6.Variability.
Predictability and Chaos EPS and Probability Forecasting.
2012: Hurricane Sandy 125 dead, 60+ billion dollars damage.
Validation and Monitoring Measures of Accuracy Combining Forecasts Managing the Forecasting Process Monitoring & Control.
Regression Analysis. Unscheduled Maintenance Issue: l 36 flight squadrons l Each experiences unscheduled maintenance actions (UMAs) l UMAs costs $1000.
Chapter 12 - Forecasting Forecasting is important in the business decision-making process in which a current choice or decision has future implications:
Forecasting.
Reliability Trends of the Global Forecast System Model Output Statistical Guidance in the Northeastern U.S. A Statistical Analysis with Operational Forecasting.
MOS Developed by and Run at the NWS Meteorological Development Lab (MDL) Full range of products available at:
Hydrometeorological Prediction Center HPC Medium Range Grid Improvements Mike Schichtel, Chris Bailey, Keith Brill, and David Novak.
Transitioning unique NASA data and research technologies to the NWS 1 Evaluation of WRF Using High-Resolution Soil Initial Conditions from the NASA Land.
ForecastingOMS 335 Welcome to Forecasting Summer Semester 2002 Introduction.
1 Simple Linear Regression Chapter Introduction In this chapter we examine the relationship among interval variables via a mathematical equation.
Ch. 6 The Normal Distribution
MOS Performance MOS significantly improves on the skill of model output. National Weather Service verification statistics have shown a narrowing gap between.
1 Seventh Lecture Error Analysis Instrumentation and Product Testing.
1 BA 555 Practical Business Analysis Review of Statistics Confidence Interval Estimation Hypothesis Testing Linear Regression Analysis Introduction Case.
Ensemble Post-Processing and it’s Potential Benefits for the Operational Forecaster Michael Erickson and Brian A. Colle School of Marine and Atmospheric.
A Regression Model for Ensemble Forecasts David Unger Climate Prediction Center.
Applied Business Forecasting and Planning
Weather Forecasting - II. Review The forecasting of weather by high-speed computers is known as numerical weather prediction. Mathematical models that.
Slides 13b: Time-Series Models; Measuring Forecast Error
Forecasting and Numerical Weather Prediction (NWP) NOWcasting Description of atmospheric models Specific Models Types of variables and how to determine.
Chapter 3: Examining relationships between Data
STAT02 - Descriptive statistics (cont.) 1 Descriptive statistics (cont.) Lecturer: Smilen Dimitrov Applied statistics for testing and evaluation – MED4.
December 5, 2012Introduction to Artificial Intelligence Lecture 20: Neural Network Application Design III 1 Example I: Predicting the Weather Since the.
Lecture 12 Statistical Inference (Estimation) Point and Interval estimation By Aziza Munir.
Quantitative Skills 1: Graphing
An Analysis of Eta Model Forecast Soundings in Radiation Fog Forecasting Steve Amburn National Weather Service, Tulsa, OK.
Nature of Science Science Nature of Science Scientific methods Formulation of a hypothesis Formulation of a hypothesis Survey literature/Archives.
June 19, 2007 GRIDDED MOS STARTS WITH POINT (STATION) MOS STARTS WITH POINT (STATION) MOS –Essentially the same MOS that is in text bulletins –Number and.
SWFDP-Eastern Africa November 2011 NOAA/NCEP African Desk Product Surfing Presented by Hamza Kabelwa Prepared by Richard H. Grumm Contributions by Vadlamani.
Celeste Saulo and Juan Ruiz CIMA (CONICET/UBA) – DCAO (FCEN –UBA)
Model validation Simon Mason Seasonal Forecasting Using the Climate Predictability Tool Bangkok, Thailand, 12 – 16 January 2015.
Copyright © 2014, 2011 Pearson Education, Inc. 1 Chapter 19 Linear Patterns.
Measures of central tendency are statistics that express the most typical or average scores in a distribution These measures are: The Mode The Median.
MGS3100_04.ppt/Sep 29, 2015/Page 1 Georgia State University - Confidential MGS 3100 Business Analysis Regression Sep 29 and 30, 2015.
Ensemble Forecasting and You The very basics Richard H. Grumm National Weather Service State College PA
MODEL OUTPUT STATISTICS (MOS) TEMPERATURE FORECAST VERIFICATION JJA 2011 Benjamin Campbell April 24,2012 EAS 4480.
Objectives 2.1Scatterplots  Scatterplots  Explanatory and response variables  Interpreting scatterplots  Outliers Adapted from authors’ slides © 2012.
This material is approved for public release. Distribution is limited by the Software Engineering Institute to attendees. Sponsored by the U.S. Department.
Model Post Processing. Model Output Can Usually Be Improved with Post Processing Can remove systematic bias Can produce probabilistic information from.
CC Hennon ATMS 350 UNC Asheville Model Output Statistics Transforming model output into useful forecast parameters.
NCEP Models and Ensembles By Richard H. Grumm National Weather Service State College PA and Robert Hart The Pennsylvania State University.
Time Series Analysis and Forecasting. Introduction to Time Series Analysis A time-series is a set of observations on a quantitative variable collected.
Statistical Post Processing - Using Reforecast to Improve GEFS Forecast Yuejian Zhu Hong Guan and Bo Cui ECM/NCEP/NWS Dec. 3 rd 2013 Acknowledgements:
Regression Analysis: Part 2 Inference Dummies / Interactions Multicollinearity / Heteroscedasticity Residual Analysis / Outliers.
Correlation They go together like salt and pepper… like oil and vinegar… like bread and butter… etc.
The Record South Carolina Rainfall Event of 3-5 October 2015: NCEP Forecast Suite Success story John LaCorte Richard H. Grumm and Charles Ross National.
An Examination Of Interesting Properties Regarding A Physics Ensemble 2012 WRF Users’ Workshop Nick P. Bassill June 28 th, 2012.
On the Challenges of Identifying the “Best” Ensemble Member in Operational Forecasting David Bright NOAA/Storm Prediction Center Paul Nutter CIMMS/Univ.
An Ensemble Primer NCEP Ensemble Products By Richard H. Grumm National Weather Service State College PA and Paul Knight The Pennsylvania State University.
CFI GROUP WORLDWIDE ANN ARBOR ATLANTA BUENOS AIRES KUALA LUMPUR LONDON MADRID MELBOURNE MILAN PARIS PORTO ALEGRE SEOUL SHANGHAI STOCKHOLM National Weather.
DOWNSCALING GLOBAL MEDIUM RANGE METEOROLOGICAL PREDICTIONS FOR FLOOD PREDICTION Nathalie Voisin, Andy W. Wood, Dennis P. Lettenmaier University of Washington,
VERIFICATION OF A DOWNSCALING SEQUENCE APPLIED TO MEDIUM RANGE METEOROLOGICAL PREDICTIONS FOR GLOBAL FLOOD PREDICTION Nathalie Voisin, Andy W. Wood and.
Verification methods - towards a user oriented verification The verification group.
Chapter 6: Descriptive Statistics. Learning Objectives Describe statistical measures used in descriptive statistics Compute measures of central tendency.
Demand Management and Forecasting Chapter 11 Portions Copyright © 2010 by The McGraw-Hill Companies, Inc. All rights reserved. McGraw-Hill/Irwin.
Figures from “The ECMWF Ensemble Prediction System”
Chapter 11 – With Woodruff Modications Demand Management and Forecasting Copyright © 2010 by The McGraw-Hill Companies, Inc. All rights reserved.McGraw-Hill/Irwin.
of Temperature in the San Francisco Bay Area
Question 1 Given that the globe is warming, why does the DJF outlook favor below-average temperatures in the southeastern U. S.? Climate variability on.
of Temperature in the San Francisco Bay Area
Model Post Processing.
MOS Developed by and Run at the NWS Meteorological Development Lab (MDL) Full range of products available at:
Post Processing.
Rapid Adjustment of Forecast Trajectories: Improving short-term forecast skill through statistical post-processing Nina Schuhen, Thordis L. Thorarinsdottir.
Presentation transcript:

Ensembling Medium Range Forecast MOS GUIANCE By Richard H. Grumm National Weather Service State College PA and Robert Hart The Florida State University

Introduction Model Output Statistics (MOS) –Regression equations of parameters from models to make a forecast for a point. –statistical equations were specifically tailored for each location, taking into account factors such as local climate. MOS equations are based on output from a single model –Models have bias and errors –This will affect, though some statistical corrections, the MOS output –Timing errors and intensity errors will affect the outcome

Introduction-II Verification Issues –Base line the extended MOS and ENSEMBLE of extended MOS with more widely used products FWC - NGM based MOS MET – Eta based MOS MAV – GFS based MOS –Limited at this time to temperatures only Average (bias or avg ), Mean Absolute Error ( MAE ) and root-mean square error ( RMSE ). Primary Goal is to evaluate the longer term MOS guidance and an Ensemble –But the shorter term MOS helps illustrate the point –Helps us further learn about ensembles and ensemble strategies.

Medium Range Forecast MOS Produced from data from the Global forecast system (GFS) at the highest resolution of the model –Known as the Extended Range GFS MOS (MEX) –Message contains routine variables and a climate range of data at some locations at the end of the message –Most widely used extended range MOS product –Based on the highest resolution forecast model

MEX Ensemble MOS NCEP runs the GFS and has an GFS ensemble prediction system (EPS) –MOS is generated for each GFS EPS member –The control run is slightly coarser resolution than the operational GFS –The control is used to produce perturbed members There are 5 positive (P1-P5) and 5 negatively (N1-N5) perturbed members based on the control run This provides 12 GFS runs to produce MEX guidance the operational MRF MOS prediction equations are applied to the output from each of the ensemble runs

Why Ensemble MOS? The operation GFS and its MOS – are expected to be more accurate – Due to higher resolution which we expect to be more accurate – But may pay for timing and intensity errors as finer scale systems typically correlate less in space. The ENSEMBLE of the MOS…. –Will show a range and times of high disagreement ~ UNCERTAINTY

Ensembles help us quick review of ensemble concepts Deal with uncertainties in data (initial conditions) – the ability to properly resolve the feature Deal with uncertainties in data verse resolution of the model – 6  rule, we may under sample a system. Deal with uncertainties in physics – Current GFS system is based solely on the same model – Only variation is initial conditions

GFS EPS with varied initial Conditions Forecast Length Forecasts Initialized at most recent data time Envelope of solutions at single time Solution

Displaying uncertainties in forecasts In Model output: – spaghetti plots and probability charts (the most likely outcome) – consensus forecast charts, the middle ground, with dispersion (standard deviation about the mean) – to visualize these is to see limits of any single solution. MOS OUTPUT: – Unless plan view maps, does not lend itself well to spaghetti plots – In text bulletins, the dispersion about the mean may show uncertainty – Consensus and the range of possibilities are good candidates for display Why ensemble MOS output 

Consider this Consider this: Would you want to shoot one arrow at the bullseye or a quiver full of arrows at the bullseye. Or…pick the MEX and maybe miss the mark or use the ensembles and have a better chance of approximating something something.

Producing Ensemble MOS Assume all members of equal skill – Any single member may be most correct at any single time frame – No a priori knowledge as to which would be best member on any give day or forecast length Decode each product – Computer sums, sums of squares etc Text variables are assigned numbers – Clouds: CL = 0; SC: 33 BK: 66 and OV is 100 – Java Object can translate number to letter or vise/verse – Get range – Produce consensus and dispersion of forecasts about the mean. For select parameters show : – MEX value, consensus value, highest and lowest members value – Current system has no weights applied.

Extended MOS Formatting is lost from web

Short Term Ensemble MOS KAOO

Verifying Ensemble MOS At select locations compute – Mean errors (bias) – M ean a bsolute e rror (MAE), and – R oot m ean s quare e rror (RMSE) How to view data with 12 members per site and 150 sites verified for this study? Lots of graphs, currently at single sites.

Reference values CTP 1-4 day forecasts Dec 2003 MOSBIAS MAE MAV ETA FWC CCF

Short Term MOS Verification Examine MAV, ETA, FWC –06 and 18 UTC MAV provide 18,30,42,54 and 66 hour forecasts –ETA,FWC,MAV at 0000 and 1200 UTC provide 24,36,48,60 and MAV 72 hour forecasts –So, 6 hour forecast intervals for MAV Not to belabor point, but each MAV update clearly improves on previous as skill decreases with time. The false belief that the 06 and 18UTC MAV is unfounded. Select sites are shown though data exists for about 100 sites No scoring of ensemble products is accomplished here

KMDT AVG-MAE FWC cold bias MAV lower MAE

KMDT RMSE MAV lower RMSE

KIPT AVG-MAE FWC cold bias MAV lower MAE

KIPT RMSE MAV lower RMSE

KBOS AVG-MAE FWC cold bias The ETA/FWC bias are opposite ~enemble potential ? MAV slightly lower MAE

KBOS RMSE MAV lower RMSE But Eta is close

Short Term MOS findings At sites examined (3 shown): – MAV is the most skillful temperature forecast MOS – FWC has cold bias at many (most) sites – Eta MOS has some regional/local variation and is more competitive with MAV as some sites with opposite bias from FWC  this may lend itself to ensembling. – Clear skill differences at some sites where MAV is far superior  May limit ensembling without weighting at these sites as straight blend would weaken impact of more accurate member

Medium Range MOS Similar display concepts Same observational data sets used Plot all 12 members –P members are RED N members are blue –Focus on the 3 best members: MEX with finest detail forecast (GREEn) Ensemble mean (BLACK) Control Run (YELLOW)

KMDT BIAS JA-FE 2004 MEX consensus MEX-Control

KMDT MAE JA-FE 2004 Early on MEX is far superior!

KMDT RMSE JA-FE 2004

Allentown MAE

Atlantic City MAE

Boston MAE

LGA MAE

KABE MAE

BFD MAE

Some findings RMSE at hours comparable to those found in other MOS products 3 to 4 degrees – May be a bit better than MAV at 24 hours at KMDT MEX clearly outperforms all members and consensus at most sites hours – Ensembling unskillful members is not helpful? MEX and Consensus both very good at 60+ hours – MEX more skillful at some sites than consensus, but consensus more skillful at others – Consensus, treated as a single forecast member is quite a skillful member at all locations after 60 hours Weighting the consensus toward the more skillful members might improve the forecasts

Ensemble Envelope Useful to know how often the warmest and coldest members captured the range of solutions…. At mid-range forecasts –50-60% of the time the observed temperature is within the forecast range of all 12 members –60% of the time at longer ranges, the observed temperature is outside of the range of the ensemble range –Artifact of untuned MOS for GFS EPS members?

KMDT when observed Temperature within range of Ensemble MOS forecasts

KBFD when observed Temperature within range of Ensemble MOS forecasts

KBOS when observed Temperature within range of Ensemble MOS forecasts

A few more items Examining just January –The control run was slightly more skillful at many sites than the MEX –The consensus was typically more skillful than the MEX –A pattern change likely contributed to this, however over 2 months this problem was mitigated.

January Potential Case Study Cold snap – caused timing errors in MEX  CTL run benefited as did Consensus from the changes

Short Term MOS findings At sites examined (3 shown): – MAV is the most skillful temperature forecast MOS – FWC has cold bias at many (most) sites – Eta MOS has some regional/local variation and is more competitive with MAV as some sites with opposite bias from FWC  this may lend itself to ensembling. – Clear skill differences at some sites where MAV is far superior  May limit ensembling without weighting at these sites as straight blend would weaken impact of more accurate member

Conclusions Short Term MOS findings: – MAV is the most skillful temperature forecast MOS – FWC has cold bias at many (most) sites – Eta MOS has some regional/local variation and is more competitive with MAV as some sites with opposite bias from FWC – Clear skill differences at some sites where MAV is far superior Medium Range Ensemble MOS was generated – From collections MOS forecasts from Various NCEP EPS model runs And NCEP short range forecast models Each model produced had different initial conditions – ensemble mean (consensus) temperature forecasts were skillful and competitive with the MEX forecasts at all sites after 60 hours

Conclusions-II Limitations – Ideally, the ensemble MOS would beat the MEX at all sites after the initial time periods. – It does not implying over the long haul: The MEX is more skillful than the other members – Especially at 24 to 60 hours! The MEX equations are tuned to the operational GFS and are not tuned to the GFS EPS members. Tuned equations for each member might improve the MOS guidance for each member and the ensemble MOS system in general. – Observed temperatures often falls outside the ensemble envelope 60% of the time at longer ranges Is this a problems? – What is the value and what are the limitations of ensembling unequally skillful members?

Conclusions-III Operational Applications – Consensus and the dispersion about the mean show times of large uncertainty – Forecasters need to apply knowledge of this uncertainty in forecasts – And this information needs to be conveyed to the users of these forecasts – Times of uncertainty are times of the ensemble providing more value to the forecast process. Plans – Apply same technique to verify the POP forecasts from these data – Experiment with weights to improve the consensus forecasts. – Improve the verification methods and software

NCEP EPS Breeding N SEEDS GIVE 2*N PERTURBATIONS Scaled + perturbation Initial random seed Opposite sign is negative perturbation Adjust magnitude to typical analysis errors 12-h forecast CONTROL-CTL Complete cycle forecast