Presentation on theme: "1 (B1) EPS design, objectives and interpretation 2009.05.12-15 1st TRCG Technical Forum Takuya KOMORI ( ) Numerical Prediction."— Presentation transcript:
1 (B1) EPS design, objectives and interpretation st TRCG Technical Forum Takuya KOMORI ( ) Numerical Prediction Division Japan Meteorological Agency “Why we need Ensemble Prediction System for TC forecasts?”
2 (B1) EPS design, objectives and interpretation “Why we need Ensemble Prediction System for TC forecasts?” (C2) EPS perspectives “Status of TIGGE Activities of JMA” (B2) Application for TC forecasts “How to use the products of JMA Ensemble Prediction System?” My Lectures in 1 st TRCG Technical Forum 12 May (Tue) 13 May (Wed) 14 May (Thu) “Please feel free to ask me a question at any time in the lectures.”
3 1.Introduction of Ensemble Forecast Objectives Statistical Methods 2.JMA Numerical Weather Prediction System GSM Forecast Performance Typhoon Ensemble Prediction System (TEPS) One-week Ensemble Prediction System (WEPS) Updated Typhoon Bogus System in March Introduction of JMA EPS-WEB Contents
4 Errors in Initial Condition –Errors in Raw Observational Data –Errors in Objective Analysis Procedures for NWP Model Errors in Forecast Model (NWP model) –Limitation in the Spatial Resolution –Errors in Physical Processes Owing to the strong non-linearity and chaotic nature of atmosphere, the small error grows with forecast time immediately. The forecast error is inevitable for Numerical Weather Prediction (NWP). There are two sources of error in NWP system; Forecast Uncertainty: Sources of Error
5 All forecasted typhoon tracks (for Typhoon Dolphin, T0822) of JMA Operational Global Spectral Model (green Lines). If we know that the TC position error of GSM forecast will be large in advance, we can use the forecast with care. Time Series of Forecast Error Size: TC Position Typhoon Dolphin took a sharp turn in the tropics. At that time, GSM could not swim with the Dolphin … Larger Position Error Best track of Typhoon Dolphin (black line) The forecast position errors are not constant and change from day to day. The growth rate of forecast error depends on the atmospheric condition. The ensemble forecasting aims at estimating the rage of forecast error in advance.
6 Analysis Field Single Deterministic Forecast There is no information about uncertainty of forecast. Single deterministic forecast Basic Numerical Weather Prediction System provide “one future atmospheric condition” starting from one initial atmospheric condition. truth Forecast Error Estimate of Forecast Uncertainty
7 In addition to the future atmospheric condition, we would like to predict the range of forecast error: Probability Density Function (PDF) of atmospheric condition. Initial PDF of atmospheric condition Analysis Field Single Deterministic Forecast Prediction of PDF Single deterministic forecast (control run) Target is to know this size and shape It is difficult to estimate the PDF, directly. The Ensemble Prediction System is a feasible method to estimate the PDF of forecast field. truth Forecast Error Forecasted PDF of atmospheric condition Estimate of Forecast Uncertainty
8 The ensemble prediction is a set of forecasts generated by NWP started from only slightly different initial conditions (added the perturbation into the analysis field). Initial Time x(t0) Forecast x(t 1 ) Forecast P(t=t 2 ) Forecast of Each Member Forecast by same model Analysis Field Deterministic Forecast Ensemble Mean Initial PDF is represented by multiple initial conditions which are slightly different from analysis field. Future PDF (range of forecast error ) is estimated from multiple forecast. Schematic Image of the Ensemble Prediction
9 JMA One-Week Ensemble Prediction System (WEPS) provides much information (51 forecast fields). JMA Typhoon Ensemble Prediction System (TEPS) has 11 members. The statistical method is useful to compact and overview the ensemble forecast information. Ensemble mean, Ensemble Spread, etc… Forecasts of Ensemble Prediction System
10 “Ensemble mean forecast” is average of all ensemble forecasts. It is approximately equal to the central value of forecast PDF. Statistically, ensemble mean error, which is defined as the mean square distance between ensemble mean forecast and analysis, is almost half of deterministic forecast’s error. Ensemble Mean Forecast Analysis Field Deterministic Forecast Ensemble Mean Forecast
11 Several scenarios are included in the ensemble forecast with large spread. Small spread of ensemble forecast represents one scenario as deterministic forecast. Initial Condition Large Spread Small Spread Ensemble Spread Spread is the standard deviation of PDF. Generally, the “ensemble spread” is defined as the average of the root-mean-square distance between the ensemble mean forecast and each ensemble forecast.
12 JMA Typhoon EPS: Case Study Small Spread (T0815 Jangmi) Typhoon EPS captures the best track with its large spread. Large Spread (T0822 Dolphin) The Ensemble Prediction System is a feasible method to estimate the range of forecast error (PDF of forecast field) in advance.
13 JMA Numerical Weather Prediction System
14 Deterministic Models (March 2009) 20,000km 2,000km 200km 20km 2km 200m Global Spectral Model (GSM) Global Spectral Model (GSM) Meso-Scale Model (MSM) planetary wave cloud cluster Thunder storm cumulo- nimbus Spatial Scale 0.1 hour1 hour10 hour1 day1 week100 hour Temporal Scale tornado Baroclinic wave Macro Meso Micro extratropical low tropical cyclone front system
15 Ensemble Prediction Systems (March 2009) 20,000km 2,000km 200km 20km 2km 200m One-Week EPS 51 members Once a day planetary wave extratropical low cloud cluster Thunder storm cumulo- nimbus Spatial Scale 0.1 hour1 hour10 hour1 day1 week100 hour Temporal Scale tornado Baroclinic wave Macro Meso Micro front system Typhoon EPS 11 members 4 times a day
16 Specifications of the NWP Models at JMA (Mar. 2009) In addition, One-month EPS and Seasonal EPS have been in operational.
17 RMSE of 500 hPa Geopotential Height in Northern Hemisphere (GSM: T+48h) T213L30 Revision of cumulus parameterization T213L40 Revision of cumulus parameterization Revision of cloud Use of MODIS Revision of radiation TL319L40 4D-Var Use of SSM/I and TMI Variational bias correction Use of QuikSCAT Direct use of ATOVS 3D-Var Reduced Gaussian Grid TL959L60
18 5-day forecasts 3-day forecasts 4-day forecasts 2-day forecasts 1-day forecasts The position error of the 5-day forecasts in 2007 is equal to that of the 3-day forecasts in The position error of the 5-day forecasts in 2007 is equal to that of the 3-day forecasts in Verification of 5-day TC Track Forecasts by GSM
19 Typhoon KROSA: 6-7 October 2007 Typhoon KROSA simulation by 20kmGSM Analyzed Track by JMA GSM precisely simulated the rotate motion of KROSA.
20 MTSAT-IR Observation GSM Forecast Typhoon JANGMI: 25 Sep GSM provided good forecast performance for recurvature of Typhoon JANGMI.
21 “Nevertheless, do we need Ensemble Prediction Systems for TC Forecasts?”
22 Position error (km) average Position errors of each TC Track Forecast by GSM Forecast time: 72 hours Total number of forecast events: 102 Position errors are sorted in ascending order Why is a probabilistic approach needed?
23 Specification of JMA Typhoon EPS (TEPS) Period ~ ModelLow-resolution GSM Horizontal Grid Vertical Layer60 Member11 Forecast time (h)132 Perturbation MethodSV The initial condition for the control forecast of TEPS is generated by simply eliminating higher wave number components of the high resolution global analysis. The ensemble perturbations are generated by the singular vector (SV) method and added to the control initial condition.
24 SV Target Areas for TEPS Perturbation is calculated by the Singular Vector (SV) method in adaptive areas surrounding each tropical cyclone and an additional fixed area of 20-60N/ E.
25 Generated Perturbations for TEPS One of the Generated Perturbations for TEPS (Wind Component) Singular Vector Method Component of the state vector : T, Q, Vor, Div, Psurf
26 TEPS Case Study: UTC （ T0807 KALMAEGI ） TEPS TL319L60 (FT132) GSM TL959L60 (FT84) Typhoon EPS could capture the best track with its large spread.
27 TEPS TL319L60 (FT132) GSM TL959L60 (FT84) Spread of Typhoon EPS changes depending on the forecast uncertainty. TEPS Case Study: UTC （ T0815 JANGMI ） You can get more information about GSM (Nakagawa, 2009) and Typhoon EPS (Yamaguchi and Komori, 2009) via the following website of RSMC Tokyo – Typhoon Center;
28 Specification of JMA One-Week EPS (WEPS) Period ~ ~ ~ ~ Horizontal Grid (deg.) Vertical Layer40 60 Member25 51 Forecast time (h)216 Perturbation method BGM SV Perturbation areaNorth of 20NNorth of 20S = 51 global predictions for one ensemble forecast starting at one initial time. … Both TEPS and WEPS use same model (low-resolution version of GSM) and analysis field. The One-week EPS consists of 1 control-run which forecasts from low- resolution analysis field, and 50 perturbed-runs.
29 Control analysis Perturbation The initial conditions of ensemble member are defined by adding (subtracting) initial perturbation to the control analysis field. Initial perturbation is generated by Singular Vectors (SVs). Perturbed analysis = Uncertainty of initial condition Initial Perturbation for WEPS 25 x 2 51 ( x 2)
30 Specifications of TEPS and WEPS at JMA Yamaguchi and Komori (2009);
31 Probability of Precipitation Reliability of Forecast (represented as A, B or C) Error Range of High/Low Temperature JMA issues the probabilistic forecast for public users Probability and uncertainty used in the One-week forecast is derived from the One-week Ensemble Prediction System (WEPS). One-Week Forecast Website for public users at JMA
32 Recent Revision of Typhoon Bogus Technique (Mar. 2009)
33 For Preventing disasters –It is important to analyze TC correctly for exact forecast. –Location, size, intensity, … –In many cases, observational data do not exist around a TC to represent its structure properly (TC locates in the ocean). It must be always high quality. –Directly-connected to heavy disasters. It is operational, not research. –Although the basis is physically consistent, technical adjustments are sometimes required. Operational TC Bogus - Necessity -
34 GSM –T L 959L60 –20km res. in horizontal with top level at 0.1hPa Typhoon EPS –Using GSM –T L 319L60, 11 members One-week EPS –Using GSM –T L 319L60, 51 members Regional model MSM (Meso Scale Model) –NHM (Non-Hydrostatic Model) –5km res. in horizontal with 50 layers in vertical. –4D-var (res. 10km outer, 20km inner, hydrostatic) Models using TC Bogus Data The pseudo observational bogus data are put into analysis as upper observational data (1000,925,850,800,700,600,500,400hPa ).
35 Revision of Typhoon Bogus Technique OLD Bogus Technique NEW Bogus Technique Horizontal distribution of the pseudo data is revised on the basis of statistical investigation. First guess field (Vor. 850hPa) Analysis field (Vor. 850hPa) Bogus data (wind vector 850hPa) First guess (wind vector 850hPa) Analysis (wind vector 850hPa) TC central position in the first guess field is properly fit to the analyzed position.
36 Let’s take a break now, and resume in 5 minutes.
37 Operational forecast RSMC TC Advisory -for 3-day forecast (WTPQ20-25 RJTD: via GTS) RSMC TC Advisory -for 5-day track forecast (WTPQ50-55 RJTD: via GTS) Prognostic reasoning (WTPQ30-35 RJTD: via GTS) RSMC Tokyo-Typhoon Center Website -for 3-day forecast & 5-day track forecast (open) Numerical Weather Prediction RSMC Guidance for Forecast -TC prediction of JMA/GSM -TC prediction of JMA/GSM -TC prediction of JMA/Typhoon EPS (FXPQ20-25 RJTD: via GTS) Numerical Typhoon Prediction Web Site -TC prediction of major NWP centers in the world (for registered users) JMA EPS-WEB -JMA one-week EPS products (for registered users) RSMC Data Serving System -GPV data of global atmosphere, global wave (for registered users) Observation & analysis SAREP -tropical cyclone (TC) satellite image analysis (TCNA20/21 RJTD, IUCC10 RJTD: via GTS) RSMC Data Serving System –surface, upper air, atmospheric motion vector (for registered users) GMSLPD -satellite image analysis tool Real-time JMA Products and Tools for TC operations
38 Ensemble products on the “JMA EPS-WEB”
39 The products in JMA EPS-WEB are recommended by Manual on the GDPFS (WMO No.485). In addition to the web-site for public users, JMA provides a web-site for meteorologists and forecasters in foreign countries. The special forecast products derived from WEPS are disseminated on the website, “JMA EPS-WEB”, supporting the activity of National Meteorological and Hydrological Services (NMHSs) in Asia. The data in this website is available for operational weather forecasting in your countries. JMA EPS-WEB
40 JMA operates an EPS web-site (EPS-WEB) for supporting the activity of National Meteorological and Hydrological Services (NMHSs). The EPS-WEB is intended for NMHSs forecasters, not for public use. This web site provides the JMA One-week EPS products. Caution! The links to this website are strictly prohibited. Address of this web site is …. JMA EPS-WEB provides visualized EPS products. Introduction (JMA EPS-WEB)
41 JMA EPS-WEB (Visualized EPS products) This website provides products derived from the Ensemble Prediction System (EPS) of Japan Meteorological Agency (JMA) to National Meteorological and Hydrological Services (NMHSs), as part of a pilot project of JMA aiming at improving the EPS and increasing the availability of its products. For further improvement of JMA's EPS and its products, users' feedback, especially on comparison of EPS products with local observation and actual weather, is welcomed and should be directed to the JMA Medium-range EPS group at the following address: Information on this website is guidance and intended to be provided for use by forecasters and meteorologists in NMHSs. Please consult your national meteorological agency/administration for official forecasts of your country, region and/or local area.
42 All contents are updated at 03 UTC JMA EPS-WEB (Visualized EPS products)
JMA EPS-WEB (Visualized EPS products)
44 Contents 1: Forecast Chart (Stamp Map)
45 “Stamp” displays all synoptic weather forecast maps, 51 maps with ensemble mean, ensemble spread and spaghetti map derived from WEPS, at the same valid time. The range of forecast time is from initial to 9-day with 12-hour interval. mean spread spaghetti 50 perturbed runs Control “Forecast Chart” – all members -
46 “Sequence” displays a forecast map of selected member from initial time up to 9-day forecast. FT=0FT=12hrFT=24hrFT=36hrFT=48hr FT=60hr FT=120hr FT=72hrFT=84hrFT=96hrFT=108hr FT=132hrFT=144hrFT=156hrFT=168hr FT=180hrFT=192hrFT=204hrFT=216hr “Forecast Chart” – each forecast -
47 500hPa Geopotential Height Mean Sea Level Pressure Asia WesternPacificSouthChinaSea “Forecast Chart” – area and element -
48 Spread Spaghetti diagram Control RunEnsemble Mean “Forecast chart “ page consists of the following five charts. Ensemble Member Configurations of “Forecast Chart”
49 Control Run Control Run is a forecast from an unperturbed initial condition (analysis field). Ensemble Member Ensemble member is a forecast from perturbed initial conditions (slightly different initial conditions from analysis field ). Control Run and Ensemble Members
50 The ensemble mean is an average of all ensemble forecasts (50-perturbed and 1-control run). ex. The Ensemble mean of geopotential height at 500hPa ……… 51-ensmble forecasts averaging Ensemble Mean – definition -
51 Advantages ‘Ensemble Mean’ takes entire the data sample into account. For normal distribution, ‘Ensemble Mean’ is the most reliable measure to infer the central tendency by using multiple sample data (ensemble member forecasts). Disadvantages Not representative of the central tendency when the data sample is skewed or multi-modal distribution. The forecast fields can be inappropriately smoothed. Ensemble Mean – characteristics -
52 Normal Distribution The simple average of ensemble members is the best estimation of most likely forecast. Multi-modal Distribution The simple average of ensemble members is not always the best estimation of most likely forecast. Additional information is required to interpret the ensemble forecast. Distribution of EPS forecasts mean Forecast variable Ensemble Mean – member distribution –
53 Limitation on the usage of Ensemble Mean The ensemble mean makes sense only when a set of ensemble members show a 'Normal Distribution'. Furthermore, this ensemble mean should have the maximum likelihood of occurrence among those ensemble members. If ensemble members are clustered to two or more possible distributions, the ensemble mean (as the most likely solution) can be misled. Interpretation of the ensemble mean, especially when there are multiple likely forecast, requires additional information about the distribution of individual ensemble member. Ensemble Mean – caution –
54 Ensemble spread field at 500 hPa geopotential height. The ensemble spread measures the differences between the members in the ensemble forecast. A large spread indicates high forecast uncertainty. Ensemble Spread Map
55 The forecast of trough is different among the ensemble members. Compared to the ensemble mean, the large spread region corresponds to the vicinity of the trough. Ensemble spread Ensemble mean Ensemble Mean and Spread
56 Ex. Synoptic chart of 500hPa geopotential height. “How do we interpret the ensemble mean and ensemble spread ?” From the left chart, we recognize that the existence of the trough at 500hPa geopotential height. From ensemble mean and ensemble spread, we can estimate the uncertainty of location and existence of troughs Deterministic or Ensemble mean forecast. Trough at 500hPa geopotential height The typical spread patterns are shown in the following slides. Ensemble Mean and Spread
57 Ensemble mean chart of 500hPa geopotential height with spread. In case of a large spread with the ensemble mean trough, there is remarkable uncertainty in the amplitude of the trough. Large amplitudeSmall amplitude The trough located at the same latitude (position), but amplitude of the trough is different between ensemble members. Ensemble Mean and Spread – trough 1 -
58 Ensemble mean chart of 500hPa geopotential height with spread. In case of a large spread at one side from the ensemble mean trough (asymmetric distribution), there is a small cluster which predict different synoptic patterns. Fast-moving and deep trough An additional short trough ahead of main trough. Ensemble Mean and Spread – trough 2 -
59 When there is a large ensemble spread in the vicinity of TC, the uncertainty of TC intensity is high. Ensemble mean chart of 500hPa geopotential height with spread. Ex) Tropical Cyclone (TC) forecast (1) Ensemble Mean and Spread – low pressure 1 -
60 The control run (the lower chart) clearly indicates a strong TC, while the ensemble mean (the upper chart) indicates large and weak low areas in wide spread. In the case of TC forecast, the ensemble mean includes high uncertainties in TC position and its intensity. It is because TC low pressure can be smoothed or cancelled by averaging the ensemble forecasts. Ensemble mean chart of 500hPa geopotential height with spread. Control-run chart of 500hPa geopotential height Ex) Tropical Cyclone (TC) forecast (2) Ensemble Mean and Spread – low pressure 2 -
61 Spaghetti Diagram of geopotential height at 500hPa Spaghetti diagrams plot two or three specific contours for easy readability (5520, 5700 and 5880m line in the left diagram). It is EPS distributions at 500hPa geopotential height. FT=96 Spaghetti Diagram
62 The following charts show the spaghetti diagram of forecast at the same time, but started at different initial time. Initial time: UTC The shape of 5700m contour of ensemble member s are almost same as blue line. Initial time: UTC Several shapes of 5700m contour of the ensemble member s are superimposed. Spaghetti Diagram
63 In this case, the position of trough and ridge are different between ensemble members. FT=96 Spaghetti Diagram (sample)
64 There are three synoptic patterns in ensemble members as follows; 1. Deep trough or trough separated from flow. 2. Fast moving trough. 3. Weak (or plain) trough. 123 FT=96 Spaghetti Diagram (decomposition)
65 Thank you for your kind attention. “Questions or Comments?”