Recent advances in Tropical Cyclone prediction using ensembles Richard Swinbank, with thanks to Many colleagues in Met Office, GIFS-TIGGE WG & others HC-35 meeting, Curacao, April 2013
Recent advances in TC prediction using ensembles Ensemble prediction Introduction Multi-model ensembles Verification of TC track forecasts Forecasting Hurricane Sandy Comparing ensembles & deterministic forecasts GIFS-TIGGE Developing TC forecast products Supporting SWFDP and other FDPs & RDPs Prospects for improving intensity forecasts
Ensemble Forecasting The aim of ensemble prediction system is to represent the uncertainty in the state of the NWP model at all stages through the forecast range Each ensemble member is designed to sample a PDF representing uncertainties in the model state. The ensemble usually comprises a control run plus many perturbed forecasts. Initial condition perturbations are designed to represent uncertainties in the initial analysis closely linked to data assimilation. These perturbations grow with time as a result of the chaotic nature of model dynamics. The uncertainty in forecasts also grows as a result of model error. This effect may be represented by perturbing model physics e.g. stochastic physics schemes.
Ensemble forecasting Deterministic Forecast Initial Condition Uncertainty X Analysis Climatology time Forecast uncertainty
Examples of comparison of ensemble spread & error (from 2010 MOGREPS upgrade) Compare Red old Blue new Note: Ensemble spread is less than RMS error New system both reduces RMS error & increases spread Under-spread in surface temperature exaggerated because of representivity errors (point observations not representative of grid-square averages)
Under-dispersive forecast Initial Condition Uncertainty Real forecast uncertainty Estimated forecast uncertainty Deterministic Forecast X Analysis The ensemble may capture reality less often than it should. Dangerous - false sense of security! time
Met Office Multi-Model Ensemble One approach to improving the calibration (addressing under-dispersion) is to use multimodel ensembles (or grand ensembles). ECMWF 51 Member NCEP 21 Member Met Office 24 Member 3 variables: MSLP 2m Temp 500mb Height Met Office MME results courtesy Christine Johnson
Brier Skill Scores Probability of 2m temperature greater than climatological mean. Multi-model gives improvement in reliability and resolution at all lead times. 1 day Brier 5 day Reliability Resolution
Multi-model ensemble forecast Initial Condition Uncertainty real forecast uncertainty Estimated forecast uncertainty Deterministic Forecast X Analysis Use of multi-model ensembles can improve the sampling of forecast uncertainty time
Multi-model ensemble forecasts of T850 Demonstrates benefit of multi-model ensemble, provided that the most skilful models are used. Renate Hagedorn, ECMWF
Verification result of TC strike probability -1- Strike prob. is computed at every 1 deg. over the responsibility area of RSMC Tokyo - Typhoon Center (0-60 N, 100 E-180 ) based on the same definition as Van der Grijn (2002). Then the reliability of the probabilistic forecasts is verified. Reliability Diagram -Verification for ECMWF EPS- In an ideal system, the red line is equal to a line with a slope of 1 (black dot line). The number of samples (grid points) predicting the event is shown by dashed blue boxes, and the number of samples that the event actually happened is shown by dashed green boxes, corresponding to y axis on the right. Thanks to Munehiko Yamaguchi, MRI/JMA
Verification result of TC strike probability -2- All SMEs are over-confident (forecasted probability is larger than observed frequency), especially in the high-probability range.
Benefit of Multi-model Grand Ensemble Combine 3 single model ensembles Reliability is improved, especially in the high-probability range. MCGE reduces the missed forecasts (see green dash box at a probability of 0 %).
Benefit of Multi-model Grand Ensemble Best SME (ECMWF) MCGE-3 (ECMWF+JMA+UKMO) MCGEs reduce the missing area (missed forecasts)! The area is reduced by about 1/10 compared with the best SME. Thus the MCGEs would be more beneficial than the SMEs for those who need to avoid missing TCs or assume the worst-case scenario. MCGE-6 (CMA+CMC+ECMWF+JMA+NCEP+UKMO) MCGE-9 (All 9 SMEs)
Verification of ensemble spread Verification at 3 day predictions x axis: ensemble spread y axis: position error of ensemble mean track prediction
Verification of confidence information Position errors (km) of 1 to 5 day ensemble mean TC track predictions with small (blue), medium (orange) and large (red) ensemble spread. Each color has five filled bars, corresponding to the position errors of 1 to 5 day predictions from left to right. If If a SME is is successful in extracting the TC track confidence information, the average position error of small-spread cases is is smaller than that of mediumspread cases, and in turn smaller than the average position error of largespread cases. The frequency of each category is is set to 40%, 40% and 20%, respectively (Yamaguchi et et al. 2009).
Summary - EPS & multi-model ensembles Ensembles are valuable for forecasting the risks of exceeding thresholds (e.g. for high-impact weather events). But forecasts often need calibration to correct both biases and variability (e.g. to correct under-estimates of forecast spread). The best approach is to address the systematic errors, i.e., reduce model biases and improve the representation of model errors in the EPSs. In the mean time. Use of multi-model ensembles is a pragmatic approach that reduces calibration errors, especially where models have similar skill but different types of systematic error For tropical cyclone track predictions: Use of multi-model ensembles improves the reliability of strike probabilities The ensemble spread of track predictions gives an indication of the confidence in the ensemble mean prediction.
INTERLUDE Tropical cyclone products: examples from Met Office Uses Julian Heming s code to identify and track tropical cyclones (TC), originally developed for the deterministic global Unified Model. Tropical Cyclones (TCs) are identified where 850hPa relative vorticity (RV) maxima are greater than a threshold For TC tracking, use search radius of 4 degrees for analysis and 5 degrees for forecast positions Identified storms that do not match with a named storm or a TC identified at a previous time are counted as TC genesis Met Office TC products courtesy Piers Buchanan, Helen Titley and Julian Heming
Products for a named storm MOGREPS-15: Ma-On, 12Z Fri 14 th July 2011 Left hand plot: 24 ensemble tracks Middle plot: strike probability i.e., probability that the storm will be within 75 miles within the next 15 days. Right hand plot: MOGREPS ensemble mean (blue), control (cyan), previous observations (red) and deterministic track (green)
Hurricane Sandy - Overview Formed in southern Caribbean Sea Landfalls over Jamaica, Cuba and Bahamas Initial peak intensity 110mph 957mb near Cuba top end Cat 2 Landfall over New Jersey with intensity 85mph 946mb Peak storm surge 14.38 (4.38m) Most Sandy slides thanks to Julian Heming
ECMWF forecasts of Sandy
Met Office Deterministic Global Model Forecasts Formation predicted several days in advance Track through Caribbean and into open Atlantic very well predicted Right-of-track bias started near northern Bahamas Slow and right-of-track bias for USA landfall 00Z forecasts 12Z forecasts
MOGREPS Forecasts As Sandy was forming (7 days before US landfall) the strong probability of US landfall was flagged by MOGREPS Highest strike probabilities north of actual track (shown in black)
MOGREPS Forecasts 4 days before US landfall MOGREPS gave a good indication of landfall location
MOGREPS v. Met Office deterministic MOGREPS control and ensemble mean (blue tracks) were well to the left of the global deterministic track (green) On 23 rd & 24 th October MOGREPS gave a much better indication of US landfall. Actual track in red.
GFS ensembles 23 rd -25 th October 2012
Track comparisons from 00Z 24 th October 2012 ECMWF deterministic (dark blue) and MOGREPS ensemble mean (green) good guidance Met Office deterministic (red) and NHC guidance (light blue) much poorer
Track forecast errors Over the whole lifetime of Sandy, MOGREPS ensemble mean errors much lower than MO global deterministic model errors at longer lead times (> 72 hours)
Summary of forecasts for Sandy Met Office deterministic runs from 23 rd all had US landfall, but had a right bias which started over Bahamas ECMWF deterministic all had US landfall from 23 rd with smaller bias than Met Office model GFS deterministic did not feature a strong left turn until 12Z 25 th Nearly all MOGREPS members had US landfall from 00Z 23 rd onwards MOGREPS ensemble mean errors for whole storm lifetime lower than deterministic from T+72 onwards ECMWF ensemble flagged possibility of US landfall over 9 days ahead Majority of GFS ensemble members did not have US landfall until 24th
Towards the Global Interactive Forecast System (GIFS) Many weather forecast situations are low probability but high risk unlikely but potentially catastrophic. Probabilistic forecasting is a powerful tool to improve early warning of high-impact events. The objective of the GIFS is to realise the benefits of THORPEX research by developing and evaluating probabilistic products to deliver improved forecasts of high-impact weather. GIFS-TIGGE WG has initiated a GIFS development project to develop & evaluate products, focused on Tropical cyclones Heavy precipitation Strong winds
Tropical cyclones As a first step, the GIFS-TIGGE working group set up a pilot project for the exchange of real-time tropical cyclone predictions using Cyclone XML format. This was initiated to support the TPARC-08 campaign and has been continued. Example of combined TC track forecasts (Met Office + ECMWF)
CXML track exchange
Tropical cyclone products from MRI/JMA http://tparc.mri-jma.go.jp/cyclone/
Indian Ocean website at IMD
Use of MOGREPS-15 products in SWFDDP
Additional tropical cyclone products With the help of the EUfunded GEOWOW project, we have developed new products and are now delivering them to some SWFDP regional subprojects. Example - time series of forecast measures of cyclone intensity
Multi-model ensemble forecasts of Sandy Another new set of products is a multi-model version of the track and strike probability plots also now available for SWFDP.
How well can we forecast TC intensity? Category scores based on Vmax Global models Regional models Statistical Scores relative to statistical model using climatology & persistence Current NWP models do not add value to operational shortrange forecasts of TC intensity, but can improve on statistical models from ~day 2 onwards, when skill is very poor. Based on tropical cyclones for NW Pacific during 2010 & 2011 Hui Yu et al (WAF, early online release)
Improving Tropical Cyclone forecasts Potential for improved prediction of structure & intensity using high resolution nested ensembles. High-resolution simulation, by Stu Webster (Met Office)
Summary Ensemble forecasts are invaluable for forecasting the risks of severe weather, such as tropical cyclones. Combining forecasts from several skilful ensembles often produces better calibrated, and more statistically reliable results. The benefit of ensemble prediction has been demonstrated in forecasts of hurricane Sandy. Using TIGGE data, products are being developed to support the forecasting of tropical cyclones and other severe weather. The products will be evaluated in conjunction with the SWFDP. While current NWP models can do a good job at predicting tropical cyclone tracks, higher resolution models will probably be needed for improved intensity forecasts.