(tim.hewson@ecmwf.int) (fernando.prates@ecmwf.int) Introduction to the ECMWF products with special emphasis on using EPS for severe weather Tim Hewson (tim.hewson@ecmwf.int) Fernando Prates (fernando.prates@ecmwf.int) 1. The Extreme Forecast Index (EFI) 2. Extra-Tropical Cyclones products 3. Examples (Lab Session)

The operational forecast system High resolution deterministic forecast: twice per day 16 km 91-level, to 10 days ahead Ensemble forecast (EPS): twice daily 51 members, 32/65 km 62-level, to 15 days ahead extended to 32 days once a week (Monthly forecast) Ocean waves: twice daily Global: 10 days ahead at 28 km Limited Area Wave (LAW): 5 days ahead at 10 km Ensemble: 15 days ahead at 55 km Seasonal forecast: once a month 41-members, 125 km 62 levels, to 7 months ahead a sub-set ensemble of 11 members is run for 13 months every quarter

Example plot of the “Extreme Forecast Index” = EFI Windy Cold conditions EPS EFI 24@00 +108h VT: 28 August 2011 at 12UTC

1. Computing the Extreme Forecast Index (EFI) The EFI is NOT a probability event (rain, wind speed, temperature, etc) ….But is a measure of the difference between the ensemble forecast distribution and the M-Climate distribution

How do CDFs and PDFs relate ? PDF (probability density function) Prob (interval) Parameter value Prob (threshold) Parameter value The PDF (y-axis) value equals the slope of the CDF Steeper CDF = narrower PDF = higher confidence A step in the CDF means a bimodal PDF

Normalization factor to keep 1 Normalization factor to keep -1≤ EFI ≤ 1 Represented by pink lines below More weight to extremes of M-climate EFI does not account how much far the values are from M-Climate maxima M-climate EPS Cumulative 1 p Ff (p) 0.5 (rank/n) = EFI -1  EFI  1 *100% gives -100%  EFI  100%

1.2 The reference EPS Model Climate Why is important to have a Model Climate? To compensate for systematic model errors (increase reliability & quantify model errors); to know “how unusual today’s forecast is, relatively, to other forecast (past) events” For climate related products like the EFI, a reliable model climate is essential Ideally the model climate is a large set of EPS re-forecasts with the latest model configuration (used operationally) for a long enough period The current EPS model climate (at ECMWF) in use: Running a full EPS re-forecast suite with 4 EPS members and the Control Always for the most recent 18 years with initial conditions taken from ERA-Interim (the higher resolution successor to ERA-40). Currently runs every Thursday (therefore climate files are available only for Thursdays for 32 days, post-processed fields as for EPS (data every 6 hours). Uses latest model cycle (resolution/ physics / etc.) to allow an immediate adaptation of EFI and other model climate related products to any EPS model upgrade For more information about the model climate please refer to: http://www.ecmwf.int/products/changes/vareps-monthly/EFI_climate.html

Model Climate We now call this the ‘M-Climate’, which has a specific meaning… M-Climate means that it is a function of 4 factors: 1. Location, clearly 2. Time of year, to take account of seasonal variations 3. Forecast lead time (due to non-stationary of the forecast errors) 4. Model-version (specially, when there is a significant model upgrade, i.e, new physics parametrization scheme)

Operationally available EFI fields In the current operational system every EFI field is based on a forecast range of 24 hours or longer. Since each meteorological parameter is valid for a period the content is either an accumulated value (e.g. precipitation), a mean over a period (e.g. temperature or mean wind) or an extremum (maximum or minimum) over that period (e.g. diagnosed wind gust). Each 24-hour period variable is worked out as a post-processed value based on four 6-hourly forecast time steps. E.g. a mean over a 00-00 UTC period is a mean of the 06-12-18 and the ending 00 UTC fields. Importantly, for wind gusts, the 6 hourly wind gust values used are maxima within the preceding 6 hours (diagnosed by interrogating the model run at every time step)

Operationally available EFI fields For 2m mean temperature, 10m mean wind speed and maximum wind gust 5 forecast ranges are available 00z: T+0-24, T+24-48, T+48-72, T+72-96, T+96-120 12z: T+12-36, T+36-60, T+60-84, T+84-108, T+108-132

Operationally available EFI fields For Total precipitation 5+3 forecast ranges are available 00z: T+0-24, T+24-48, T+48-72, T+72-96, T+96-120 12z: T+12-36, T+36-60, T+60-84, T+84-108, T+108-132 00z & 12z: T+0-120, T+24-144, T+0-240

Composite ‘Anomalous Weather’ maps (Cont) - 3 EFI parameters on 1 combined chart - 2 thresholds for each weather type (0.5, 0.8) -Raw EFI plots are slightly higher resolution

Accessing New Products - The “Clickable EFI” interface All EPS grid points’ data are accessible by clicking on different ‘Anomalous Weather’ maps at different forecast ranges Different displays are possible The already available EPSgrams plus: EPSgram versions with the model climate information EPS and climate CDF curves with EFI information Etc. Click anywhere

Example: Reading on one day in June 2009 Y-axes = “probability to not exceed” 45% probability of >6mm (red line) 10% probability of >6mm on an average June day The 15-year return period 24h rainfall for ~June is 27mm (M-climate). One old forecast had indicated a small risk of reaching this, but the latest forecast (in red) didn’t. Steeper CDF slope on more recent forecasts signifies increasing confidence Downside – if parameter values are directly referenced, model biases are not accounted for

How ‘should’ CDFs behave in successive EPS runs? M-climate rank Successive EPS runs value At long leads CDF may be similar to the M-climate Lateral variations in CDF position between successive runs should, mostly, become less (with time) CDF slope will tend to increase (with time), implying higher confidence

But there are Counter Examples: obs Wind gust here Windstorm ‘Klaus’ – Jan 2009 – Atlantic point. N England rain – June ’09 - low prob alternative became likely at short range. If rare this is OK.

Precipitation CDFs for Washington DC – 6 Feb 2010 Few occasions the information content can be misleading. M-Climate on standard web products is for the 24-48h lead

{

{

{

Interpretation This line denotes a ‘1 in 100 day event’ (for the time of year), for the displayed parameter – according to the M-Climate (≈ ”once per season”)

{

2. Fronts and Extra-Tropical Cyclones 15/04/2017 2. Fronts and Extra-Tropical Cyclones 12h precipitation shown by shading – greys for mainly rain, pinks for rain-snow mix, blues for mainly snow Ranges, in mm water equivalent: 2-8mm, 8-20mm, 20-50mm, >50mm

2. Fronts and Extra-Tropical Cyclones 15/04/2017 2. Fronts and Extra-Tropical Cyclones

2. Fronts and Extra-Tropical Cyclones 15/04/2017 2. Fronts and Extra-Tropical Cyclones Help: contains a detailed description of all available products Product: lists all products Run: choose the forecast date time

Further reading …. Hewson, T.D., 1997: Objective identification of frontal wave cyclones. Meteorol. Appl., 4, 311-315. Hewson, T.D., 1998: Objective fronts. Meteorol. Appl., 5, 37-65. Hewson, T.D., 2009: Diminutive frontal waves - A link between fronts and cyclones. J. Atmos. Sci., 66, 116-132. Hewson, T.D., 2009: Tracking fronts and extra-tropical cyclones. ECMWF Newsletter, No. 121, 9-19. Hewson, T.D. & H.A. Titley, 2010: Objective identification, typing and tracking of the complete life-cycles of cyclonic features at high spatial resolution. Meteorol. Appl., 17, 355-381.

15/04/2017 Principles Forecasters make daily use of feature identification (e.g. fronts, troughs, frontal waves, lows,…), mainly because of those features are responsible for bad (or extreme) weather Nowadays manual identification of features is based largely on model data If we can identify cyclones automatically in a manner that is consistent with everyday forecasting, we will have: 1. A range of new EPS-based tools for assisting the forecaster to predict bad weather (Products) 2. A means by which bad weather can be verified, by proxy, by comparing the forecast features with features identified in model analyses (Verification)

Rationale Key aspects of our system, which make identifying the full range of features possible, are: We use a hybrid identification system, based on vorticity & mean sea level pressure. With one exception other studies have used these variables in isolation. Many of the features are required to lie on fronts (objectively defined). This accords with synoptic practice. A multi-parameter tracking scheme is employed to correctly associate features. One key aspect is the use of ‘half-time tracking’.

15/04/2017 Identification methodology is based around this conceptual model of extra- tropical cyclone development (but is not constrained by it):

Lows, Frontal Waves and Diminutive Waves Barotropic lows are simply low pressure centres Frontal waves should be well known – they represent a meeting point of cold and warm fronts (where the vorticity of the cross- front wind is positive) Diminutive waves are less well known, they represent the first sign on a synoptic chart that a frontal wave may be developing – usually signified by slight opening out of the isobars Diagnostic-graphical techniques are used to identify each type of feature

Identifying Cyclonic Features in the EPS Work on identifying features in ensembles began in about 2004 (and in operational runs in about 1996) Pressure level data provides input (T, q, u, v, Z @ 1000,925,850mb,..) Data is reprojected onto a given region at a resolution of about 50km 12-h time interval used, so that EPS-based products for can be calculated in time to be used by forecasters (Shorter interval might be better and could be used in research) Range of diagnostics computed from input data Diagnostics plotted and post-processed using a graphical package Output initially comprises ‘synoptic animations’ and simple ASCII text files showing attributes of each identified feature (‘cyclone database’)

Dalmatian Chart (mslp)

Dalmatian Chart: max 1km wind (kts), 300km radius

Dalmatian Charts - Recommended Usage Most useful between T+36 and T+192 Can sometimes be helpful at other times: At longer ranges when a persistent block is present At shorter ranges when there is large dynamical uncertainty surrounding developments (=unusually large spread) Remember to contrast the features positions (and attributes) in the control and deterministic runs, with those found in the EPS, to see how representative these single runs are Do not neglect the ‘feature type’ charts, which contain valuable information regarding the synoptic pattern

Strike Probs – Stronger Features (‘>34kts’) [1km above] Also animates

Strike Probs – Storms (‘>60kts’) Also animates

Strike Prob Charts - Recommended Usage Strike probability charts should be used: 1. At longer leads 2. Also at short leads, together with plume diagrams, whenever there is some uncertainty regarding: A) whether a cyclonic feature is splitting into two, or B) which of two closely spaced pre-existing features will develop Note that there is evidence of a small degree of skill, for the more extreme storm class, beyond 10 days

Feature Plume Diagrams These are activated by clicking on a feature on the control T+0 ‘synoptic chart’ frame Plume then displays the behaviour of that feature as tracked in the ensemble (and control and deterministic runs)…

15/04/2017 1. This is the feature the user has clicked on, along with any equivalent features in the EPS members at T+0 Those members that had an equivalent feature are listed here (0=control)

Plume Diagrams - Recommended Usage Generally use at short leads only (typically up to about T+120) Because a feature needs to be present at T+0 Beware of interpreting the ‘percentage of members tracked’ too literally Reductions in number could be due to feature splitting during the forecast Or, very occasionally, to tracking errors Note that high vorticity can sometimes be more indicative of the vigour of a system than low mslp Be aware of the potential for large forecast errors when the 300mb jet is strong (say >~150kts)

Thanks!

1.1 Overview: How can we forecast ‘extreme weather’ ? What can be used to define ‘extreme’, right across the globe ? Can we define universal thresholds ? NO Return Periods ? Effectively YES. This is highly relevant because infrastructure (transport, buildings, etc) is commonly based on the return-period concept. In part warnings relate to the resilience of that infrastructure. So we need reference threshold levels, for each weather parameter of interest, everywhere across the globe Return Periods from observational data are not available everywhere However global model forecasts provide output everywhere Therefore global model forecasts can be used to provide the climatology from which to extract Return-Period-type information Hindcasts are performed each time a model is upgraded, to provide the relevant climatology – some resolution-related imperfections, but overall a very useful dataset. This forms the basis for the EFI, or ‘Extreme Forecast Index’, whereby, for each gridpoint, the range of solutions (pdf) from the current ensemble, is compared with that gridpoint’s climatology from the hindcast data, to see how extreme these solutions are

Schematic of the EPS M-Climate (for forecasts from ~13 Mar) FEBRUARY MARCH 24 25 26 27 28 29 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 30 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 APRIL 31 1 2 3 4 5 18 -15 days +15 days 5 5 5 Goes into Day+1 climate 5 Goes into Day+6 climate 5 Sampling the available re-forecasts from ±15 day vicinity of the actual day to the Model Climate The climate sample size is: 18 years * 5 members * 5 weekly runs = 450 re-forecast fields

Example Heat wave in Europe - April 2007 Observed maximum temperature, 15th April 2007 Extreme temperatures 26-30 degrees The highest temperature on Sunday (15th of April

Heat wave in Europe - April 2007 X How about the model world ? The EPS model climatology ? What should we expect from the model ? 12-14 Mean of the model climate 25-27 18-21 95th percentile of the model climate Maximum member of the model climate

Heat wave in Europe - April 2007 High resolution forecast X Heat wave in Europe - April 2007 Unusually warm weather already 10 days in advance The location of the highest temperature was correctly forecast 6-7 days ahead The temperature was underestimated a bit even at the shortest range High resolution forecast

Heat wave - April 2007 X Day-15 forecast P(T>20 C°) P(T>Q95) EFI+SOT0.9 Anomaly P(T>Q95)

Heat wave - April 2007 X Day-11 forecast P(T>20 C°) P(T>Q95) EFI+SOT0.9 P(T>Q95) Anomaly

Heat wave - April 2007 X Day-8 forecast P(T>20 C°) P(T>Q95) EFI+SOT0.9 P(T>Q95) Anomaly

Heat wave - April 2007 X Day-6 forecast P(T>20 C°) P(T>Q95) EFI+SOT0.9 P(T>Q95) Anomaly

Heat wave - April 2007 X Day-4 forecast P(T>20 C°) P(T>Q95) EFI+SOT0.9 P(T>Q95) Anomaly

Heat wave - April 2007 X Day-2 forecast P(T>20 C°) P(T>Q95) EFI+SOT0.9 P(T>Q95) Anomaly

2. Extra-tropical cyclone products 2.1 Motivation 2.2 ????

15/04/2017 2.1 Motivation Forecasters make daily use of feature identification (e.g. fronts, troughs, frontal waves, lows,…), mainly because of those features are responsible for bad (or extreme) weather Nowadays manual identification of features is based largely on model data If we can identify cyclones automatically in a manner that is consistent with everyday forecasting, we will have: 1. A range of new EPS-based tools for assisting the forecaster to predict bad weather (Products) 2. A means by which bad weather can be verified, by proxy, by comparing the forecast features with features identified in model analyses (Verification)

Rationale In most cyclone tracking work input data is at low resolution – eg 500km In reality synoptic scale cyclones, as recognised for many years by forecasters (with good reason!), vary in scale from about 100km to over 1000km Here we set out to overcome several problems related to input data resolution, in order to identify the full range of cyclones Work has been in progress for more than 10 years. Further improvements are possible… Products Training Courses 2 Feb 2011 "Fronts & Extra-Tropical Cyclones" (tim.hewson@ecmwf.int)

Rationale (continued) Key aspects of our system, which make identifying the full range of features possible, are: We use a hybrid identification system, based on vorticity & mean sea level pressure. With one exception other studies have used these variables in isolation. Many of the features are required to lie on fronts (objectively defined). This accords with synoptic practice. A multi-parameter tracking scheme is employed to correctly associate features. One key aspect is the use of ‘half-time tracking’.

3.1 Rationale AIM TO: Make greater use of the M-Climate Allow user to visualise this data in different formats Address the problem of EPS forecasts lying beyond the M-Climate Directly compare the CDFs Provide facility to inter-compare the handling by recent runs Site specific, with clickable-map user interface

EFI distributions – overcoming limitations We can account for the ‘purple zone’ by providing a facility to display CDF graphs - anywhere in the world Additional benefits: Sight of the model climate for the user, and its extremes Sight of the full CDF for the user (within M-climate range too) Capability to intercompare recent runs’ handling, of both everyday and extreme events This moves towards providing return-period-type information As many aspects of infrastructure are based on this, should be very helpful for early warning provision As the M-climate consists of 450 realisations, the M-climate extrema correspond, approximately, to 15-year return periods (for month-long time windows)

Impact of ‘Dynamical Instability’ M-climate rank Two example CDFs parameter value If there is large uncertainty in the rate of deepening of a cyclonic feature, and/or its track, this will be reflected by a relatively shallow (mean) slope in the CDFs for affected locations (especially wind and rain, but also temperature) This ‘dynamical instability’ is relatively common in the lead up to extreme events If track uncertainty is high a step is likely in the CDF (=bimodal PDF) In all such cases the EFI values will not be especially high, because they depend on behaviour of all the members This is a weaker point of the EFI concept; forecasters must be aware of this

4. Severe Weather Examples 4.1 Heavy rainfall in the UK (October 2007) [ 4.2 Heatwave in Europe (April 2007) ]

4.1 Severe Weather Example – Heavy rainfall in the UK - 16th October 2007 Radar images, 10 UTC 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 12 UTC Radar images 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 14 UTC Radar images 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 16 UTC Radar images 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 18 UTC Radar images 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 20 UTC Radar images 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 22 UTC Radar images 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 01 UTC Radar images 15th 18 UTC – 17th 06 UTC

Heavy rainfall in the UK 16th of October 2007 Radar images, 03 UTC Radar images 15th 18 UTC – 17th 06 UTC

Analysis + Model climate EPS model climate mean

DAY + 9.5 forecast EPS mean Deterministic

DAY + 8 forecast EPS mean Deterministic

DAY + 7 forecast EPS mean Deterministic

DAY + 6 forecast EPS mean Deterministic

DAY + 5 forecast EFI Deterministic EPS mean

DAY + 4 forecast EFI Deterministic EPS mean

DAY + 3 forecast EFI Deterministic EPS mean

DAY + 2 forecast EFI Deterministic EPS mean

DAY + 1 forecast EFI Deterministic EPS mean

5. Verification EFI is rather difficult to verify ! We can however base a verification scheme on observations by linking the rarity of observed values, relative to each station’s climatological record, to the extremeness of the EFI forecast There are various ways of doing this. Many alternatives have been tested. Results presented here have assumed that an ‘extreme event’ has occurred when the 95th percentile, from the observational record, is exceeded at a given site. For site inclusion we require an observational record that is >75% complete. We then construct contingency tables for different EFI thresholds, (>0.1, >0.2, >0.3 etc) and see how many times there is an ‘extreme event’ in each such category. From the contingency tables we can extract hit rates and false alarm rates, and plot these in ROC curve format…

Example ROC Curves, Spring 2010, 24h Ppn ROC area (day 5) EFI > 0.9 EFI > 0.8 EFI >0.7 .

EFI ROC Area (covering 6 years) 15/04/2017 EFI ROC Area (covering 6 years) 10m wind speed 2m temperature (warm) 24h precipitation

6. Future Products 5.1 Pre-Operational EFI fields 5.2 EFI & M-Climate in map form 5.3 Return Periods

6.1 Pre-operational EFI fields In the near future the EFI should be extended for more parameters and forecast ranges (details below are prone to change – work in progress!) Parameters: 2m mean, minimum and maximum temperature 10m mean wind speed Maximum wind gust Total precipitation, and total snowfall Maximum significant wave height 24-hourly forecast ranges up to DAY 7: 00z & 12z: T+0-24, T+12-36, T+24-48 … T+108-132, T+120-144, T+144-168 Longer forecast ranges for characterising prolonged events which last for more days, like heat waves or wet periods : 48-hour ranges up to DAY 10 72-hour ranges up to DAY 10 120-hour ranges up to DAY 10 plus DAY 11-15 as an outlook beyond T+240 ONLY for 2m mean temperature, mean 10m wind speed and total precipitation

6.2 EFI and M-Climate in map form Snowfall (falling, not accumulated)

6.3 Return Periods Aim is to assess, statistically, the rarity of EPS forecast events, in particular those that lie beyond the M-Climate More genuinely ‘extreme’ Will use ‘extreme value theory’ to extrapolate the M-Climate data to estimate an appropriate return period Work in progress… (Fernando)

Return Period Case study – Summer 2003 heat wave VT: 18Z 12th Aug 2003 5-year 40-year Top panel: T+114h probability forecast for Tmax exceeding 5- and 40-year return period Bottom panel: Verification (hits shaded)

SUMMARY

RECAP: How can we forecast ‘extreme weather’ ? What can be used to define ‘extreme’, right across the globe ? Particular Thresholds ? NO Return Periods ? Effectively YES. This is highly relevant because infrastructure (transport, buildings, etc) is commonly based on the return-period concept. In part warnings relate to the resilience of that infrastructure. So we need reference threshold levels, for each weather parameter of interest, everywhere across the globe Return Periods from observational data are not available everywhere However global model forecasts provide output everywhere Therefore global model forecasts can be used to provide the climatology from which to extract Return-Period-type information Hindcasts are performed each time a model is upgraded, to provide the relevant climatology – some resolution-related imperfections, but overall a very useful dataset This forms the basis for the EFI, or ‘Extreme Forecast Index’, whereby, for each gridpoint, the range of solutions (pdf) from the current ensemble, is compared with that gridpoint’s climatology from the hindcast data, to see how extreme these solutions are

Some EFI limitations Extreme does not necessarily mean high impact (eg 2mm rain in a desert) Past history also important but not directly accounted for (eg heavy rain when ground saturated) Windstorm impact can depend on whether trees are in leaf, whether ground is saturated… Products are only as good as the model output, e.g.: Tropical cyclone representation is limited by resolution Threat from intense, very localised convection unlikely to be fully captured (again a resolution limitation – though with recent resolution upgrade things are slowly improving) Remember that ‘dynamical instability’ can reduce the EFI in situations where the true risk of extreme weather is relatively high Some severe weather parameters – e.g. blizzard, sandstorm, freezing rain – remain to be incorporated

Thank you ! Further Reading “ Recent developments in extreme weather forecasting” ECMWF Newsletter No. 107 (pp8-17) - Spring 2006 “Forecast Product Development at ECMWF”, Proceedings of the 11th Met Op Sys Workshop held at ECMWF, Nov 2009.