Progress in global tropical cyclone forecasting at ECMWF

Progress in global tropical cyclone forecasting at ECMWF Martin Miller Slide 1 Nargis 0600 UTC May 2 nd 1

Outline Ø Introduction § General performance improvements § Objective scores § Systematic errors Ø TC forecast skill § ECMWF: Tracks, intensity and genesis § Other Centres Ø Forecast system improvements § Physical parametrizations § Observation quality control § Resolution Ø Extended-range forecasting § Monthly § Seasonal Slide 2 Ø Conclusions 2

Historic Evolution of Skill Slide 3 3

Global extra-tropics Slide 4 4

Slide 5 5

Slide 6 6

Large-scale eg MJO Slide 7 7

ECMWF Systematic Error: D+3 DJF 1986 -1989 1996 -1999 2006 -2009 Z 500 Stream Function 200 h. Pa Velocity Potential 200 h. Pa Slide 8 8

ECMWF Systematic Error: D+3 JJA 1986 -1989 1996 -1999 2006 -2009 Z 500 Stream Function 200 h. Pa Velocity Potential 200 h. Pa Slide 9 9

Forecast system development Ø The current levels of success in TC forecasting at ECMWF are due to many important changes in the forecast system in the past, In particular some changes of note include: § implementation of 4 D-Var; § increases in horizontal and vertical model resolution; § better model physics; § availability and assimilation of massive amounts of data from satellites, and also the drop-sondes released in and around TCs; § EPS stochastic physics and also targeted perturbations in the vicinity of the TCs. § A wide-range of improvements in seasonal forecasting systems involving both Slide 16 atmosphere and ocean 16

Tropical cyclones Verification of TC predictions from the operational deterministic forecast for 12 month periods ending on 14 July. The latest period, 15 July 2008 to 14 July 2009, is shown in red. Slide 17 Mean error in core pressure (left) and position (right). 17

Tropical cyclones Verification of tropical cyclone predictions from the operational deterministic forecast for 12 -month periods ending on 14 July. Along track error: mean error in the direction of travel of the cyclone (negative values indicate slow bias) Slide 19 Cross track error: mean error at right-angles to the direction of travel 19

Slide 20 20

Nargis 0440 UTC May 1 st Slide 21 21

D+5 from 23 april 00 z Black dots are the first official reported positions of cyclone (~april 28 th) T 799 D+9 from 23 april 00 z Slide 22 22 EPS probabilities from 23 april 00 z at D+5

August-September 2008 Atlantic Basin Tropical Storm Genesis Ø Increasingly ECMWF forecasts pick up tropical cyclones before they are officially reported Ø Recent Atlantic hurricanes were predicted 5 -7 days before they were observed TS/Hurricane 1 st Observed date Forecast detection Fay 16 -Aug D+6 (run: 10 -Aug) Gustav 25 -Aug D+5 (run: 20 -Aug) Hanna 28 -Aug D+5 (run: 23 -Aug) Ike 02 -Sep D+7 (run: 26 -Aug) Josephine 03 -Sep D+5 (run: 29 -Aug) Slide 23 23

Seamless Approach to Understanding Model Error Precipitation Climatology JJA (GPCP) Analysis Increments U 925 (JJA 2008) Unit = 0. 1 ms-112 hr-1 mmday-1 -15 -9 -3 3 9 15 ms-1 day-1 DYN V. DIF CON For the India monsoon, momentum tendency is a residual of the sum of large terms. Slide 24 mmday-1 Model Climate Error JJA (32 R 3) Process 24 Tendencies U 925 (JJA 2008)

WGNE TC Verification TC tracks on 2008 season Northern-Hemisphere [2008/01/01 to 2008/12/31] Southern-Hemisphere [2007/09/01 to 2008/08/31] Number of TCs , [best track data provider] 22 western North-Pacific [RSMC Tokyo] 17 eastern North-Pacific (including Central-Pacific) [RSMC Miami, Honolulu] 16 North Atlantic [RSMC Miami] 4 north Indian-Ocean [RSMC New-Delhi] 12 south Indian-Ocean [RSMC La-Reunion] 4 around Australia [RSMC Nadi and 4 TCWCs ] 17 22 4 Slide 25 4 12 25 Daisuke Hotta and Takuya Komori (NPD/JMA) 16

(a) Verification of western North-Pacific (WNP) domain Position Error 22 TCs in 2008 Slide 26 26

(a) WNP domain Detection Rate Detection. Rate – Position. Error map (FT +72) be tte r Slide 27 27

(a) WNP domain AT-CT bias map (FT +72) JMA ECMWF UKMO CMC DWD NCEP NRL Meteo France Scattering diagram of TC positions at 72 hour forecast. Red : Before recurvature Green : During recurvature Slide 28 Blue : After recurvature Y-axis represents position errors in Along Track (AT) direction and X-axis in the Cross Track (CT) direction. Unit: kms. ECMWF has a small bias in all stages. CMC, DWD, NCEP, , NRL, Meteo France have slow bias in “before recurvature” stage. 28

(a) WNP domain Central Pressure scatter diagram (FT +72) JMA ECMWF UKMO CMC DWD NCEP NRL Meteo France Scattering diagram of central pressure at 72 hour forecast. Slide 29 Y-axis represents central pressure of forecast and X-axis does that of analysis. Unit: h. Pa 29

Time series of 2 -day and 4 -day forecast of JMA, ECM, UKM and 3 centres ensemble in WNP domain. Slide 30 30

(b) Verification of North-Atlantic (NAT) domain Position Error 16 TCs in 2008 Slide 31 31

(b) NAT domain Detection Rate Detection. Rate – Position. Error map (FT +72) be tte r Slide 32 32

A characteristic according to the domain of Northern Hemisphere. Position error of each model will be compared in the next slide Slide 33 according to the domain. 33

(d) Verification of south Indian-Ocean (SIO) domain Position Error 12 TCs in 2008 Slide 35 35

Tropical Forecast Biases Precipitation against GPCP for different cycles: from 15 year 5 months integrations for 1990 -2005. a b 31 r 1 c 32 r 2 d 32 r 3 Slide 39 39

Introduction of the Huber norm if |x| <= k, if |x| > k, Gaussian Huber Gaussian + flat Slide 40 40

The Huber-norm and robust estimation Ø 18 months of conventional data §Feb 2006 – Sep 2007 Ø Normalised fit of PDF to data § Best Gaussian fit § Best Huber norm fit Slide 41 41

Comparing optimal observation weights Huber-norm (red) vs. Gaussian+flat (blue) Ø More weight in the middle of the distribution Ø More weight on the edges of the distribution Ø More influence of data with large departures 25% §Weights: 0 – 25% Slide 42 42

Implementation of the next resolution upgrade Ø Configuration: § Deterministic model: T 1279 L 91 (~16 km) § Outer loop of 4 D-Var T 1279 L 91 and inner loops (T 159/T 255) § EPS resolution T 639 (to 10 days) and T 319 thereafter § Wave model (25 km and 36 directions) Ø Implementation planned for January 2010 Slide 44 44

OPS (35 r 2) 980 h. Pa (35 r 3) Use of improved QC (Huber norm) 961 h. Pa Hurricane Bill, observed pressure~944 h. Pa Slide 45 (36 r 1) High-res system (T 1279 etc) 45 945 h. Pa

ECMWF model simulations (T 1279, ~15 km) compared with Satellite Observations (Met 9) Slide 47 47

The MJO and tropical cyclones in the monthly forecast system ØWhat is the state of play regarding the models MJO? ØWhat is the models TC climatology like? ØHow does the MJO influence the model TC Ø 15 -member ensemble forecasts starting on the 15 th of each month from 1989 to 2008. Ø 46 -day integrations Ø Cycle 32 R 3 Slide 48 Ø T 399 uncoupled till day 10 and T 255 coupled after day 10 (Frederic Vitart – submitted to GRL) 48

MJO Propagation Time spent in each phase of the MJO Slide 49 Model 49 ERA-I

Amplitude of the MJO Slide 50 50

Tropical Cyclone Genesis climatology 1989 -2008 Observations Model NDJFMA JASON Slide 51 51

Tropical Cyclone Density climatology 1989 -2008 Observations Model NDJFMA JASON Slide 52 52

MJO Composite- ASO Tropical storm density anomaly Observations Model Phases 2+3 Phases 4+5 Phases 6+7 Phases 8+1 Slide 53 53

MJO Composite- ASO Tropical storm density anomaly Phases 6+7 – Phase 2+3 Observations Model Slide 54 54

Example of Tropical Storm Seasonal Forecast Slide 55 55

Interannual variability of Atlantic tropical storm ACE Forecasts issued in June for the period July-December 1990 Correlation: 0. 72 RMS error: 0. 40 2009 Slide 57 57

Interannual variability of the frequency of Typhoons Forecasts issued in April for the period May-October 1990 Correlation: 0. 66 RMS error: 3. 29 2009 Slide 58 58

Closing remarks Ø Steady improvements in skill in traditional measures such as 500 h. Pa geopotential (~1 day per decade for deterministic forecasts and ~1 day per 7 years for ensemble systems). Forecasts of weather parameters such as precipitation, cloud etc. are ~ 1 day per 7 years. Ø Skill of tropical cyclone forecasting is improving at a similar rate with emerging skill in intensity and genesis. ECMWF generally the best available especially in the most recent years Ø Advanced assimilation methods enable better use of storm-related data Ø Progressive increases in resolution and improved physics are improving realism of structures and intensity and hence improvements in tracking Ø Links to MJO activity provide some sub-seasonal skill in forecasting TC Slide 59 statistics Ø Seasonal forecasting shows (surprisingly? ) good 59 skill in multi-basin statistics