The Met Office GloSea5 System Description and Latest Results Jeff Knight and many other colleagues, Met Office Hadley Centre 13th Session of the Forum on Regional Climate Monitoring, Assessment and Prediction for Asia (FOCRAII), 24th April 2017
The Met Office Global Seasonal Forecast System: GloSea5
GloSea5 Met Office Global Seasonal forecast system, version 5 Model: HadGEM3 GC2 (updated Feb 2015) Resolution: Atmos, N216 L85 (~60km); Ocean: 0.25° L75 Initialisation: Daily, NWP state + NEMOVAR 0.25° Ensembles: Stochastic physics + lagged initialisation Forecasts: 2 per day -> 42 members, Hindcasts: 3 per 4 times/month -> 12 members, 1996 – 2009 Products: http://www.metoffice.gov.uk/research/climate/seasonal- to-decadal/gpc-outlooks The Korea Meteorological Administration (KMA) produce long-range forecast ensembles with a very similar system MacLachlan et al. 2014, Scaife et al. 2014 © Crown copyright Met Office
Global Coupled modelling configuration 6.0 6.0 5.0 6.0 CICE The Los Alamos Sea Ice Model N216 (~60km) ORCA025
Initialisation of the system Forecast (initialised daily): - Atmosphere & land surf *: Met Office NWP analysis (4d-Var) - Ocean & sea-ice: NEMOVAR (3d-Var joint system for ocean, med-range, monthly and seasonal) 14-year hindcast (1996-2009): - Atmosphere & land surf *: ERA-interim - Ocean & sea-ice: NEMOVAR - Fixed start dates of 1st, 9th, 17th, 25th of each month - 3 members per start date * Soil moisture set to climatological average
GloSea5 forecast schedule Seasonal Forecast: - 2 members run each day. - Seasonal forecast updated weekly by combining members from last 3 weeks (i.e. 42 members) Monthly Forecast: - 2 additional members run each day. - Monthly Forecast updated daily by combining members from last 7 days (i.e. 28 members) Hindcast (for monthly-seasonal): 14 year hindcast run in real time ( 42 members run each week = 14 years x 3 members)
A month in the life …. MacLachlan et al, 2014
Products Met Office website: http://www.metoffice.gov.uk/research/climate/sea sonal-to-decadal/gpc-outlooks Lead Centre for Long-Range Forecast Multi- Model Ensemble (hosted by Korea Meteorological Administration): https://www.wmolc.org/
Forecast maps
Timeseries (plume diagrams)
Skill scores
New developments in GloSea5
New developments Aug 2016: Mar 2017: May 2017: Increased hindcast length (1996-2009) (1993-2015) 14 years 23 years Mar 2017: Increased hindcast size 3 members 7 members, 4 times per month Effective hindcast size 12 28 May 2017: Initialised soil moisture
Impact of larger ensemble February PMSL
Initialised Soil Moisture Use Japanese Reanalysis (JRA) from the Japan Meteorological Agency: drive our land-surface model JULES in reanalysis mode update in near-real time as JRA is available in near-real time (2-day delay)
The 2015-16 El Niño: impacts in UK and China
ENSO, as we saw it in May 2015 HadISST monitoring
ENSO forecasts in 2015 Slight warming at start of 2015 From MAR Slight warming at start of 2015 Signals grew during year Very clear by Autumn From MAY From JUL From SEP From NOV
Winter 2015/16: a near record El Niño 2015/16 Forecast 2015/16 1997/8 1982/3 Very clear signals for a near record event Remote but not irrelevant Similar to 1982/83….return to this later Scaife et al, ASL, 2017
Europe: Dec 2015
North Atlantic Oscillation (single most important factor for UK winters and responds to many drivers)
Seasonal forecasting of the NAO Mild, wet and stormy Retrospective and real time forecasts from November - NAO Cold, snowy and still Ensemble Mean Observations Ensemble Member Our original tests are shown in orange and indicate a correlation skill of 62% More ensemble members => more skill and ~0.8 may be possible So far so good with real time forecasts... Scaife et al 2014, Dunstone et al 2016
December 2015 Very clear signals for a westerly winter from October From November Observations Very clear signals for a westerly winter from October Good agreement with subsequent observations Early warning of winter flooding Scaife et al, ASL, 2016
Early vs late winter and an analogue... Early winter Nov-Dec Late winter Jan-Feb 2015/16 1982/83 Toniazzo and Scaife, GRL, 2006 Fereday et al, Clim. Dyn., 2008 Remarkable similarity with 1982/3 case Remarkable similarity in late and early winter to other strong El Nino events Scaife et al, ASL, 2017
Forecasts of the polar vortex Late Winter 2015/16 Strong polar vortex Weak polar vortex Forecasts of the polar vortex NSIDC Very strong in December, weak towards late winter => low pressure and a mild, wet and stormy start to winter Scaife et al, ASL, 2017
Stratospheric conditions: winter 2015/16 A sudden warming finally happened in early March (consistent with the cold dry start to spring) Later than the most likely time in the forecasts but within the spread of forecasts from Autumn Scaife et al, ASL, 2017
Winter 2015/16: November Forecast December showed a very clear signal for wet Circulation implied increased storm risk Dec-Feb showed similar signal overall but a switch to colder in late winter Allowed real time warnings to: DEFRA, Cabinet Office and DfT
Winter 2015/16 Good agreement with subsequent observations From November Observations December Temperature December Rainfall Good agreement with subsequent observations Early warning of December flooding Driven by ENSO + few others Scaife et al, ASL, 2017
China: summer 2016
Skilful predictions of Yangtze rainfall and river flow Collaborative work with Chaofan Li and Riyu Lu (IAP), Jianglong Li (BCC-CMA) Work under the Climate Science for Services Partnership (CSSP) between the UK and China (CMA/IAP) Modest but significant grid point skill Useful regional average Skill (r = 0.55) 30 hindcast members - real time predictions have ~42 members This is an underestimate of actual forecast skill http://iopscience.iop.org/article/10.1088/1748-9326/11/9/094002/meta Li et al 2016
Larger signals than in previous forecast systems There is a larger ensemble mean signal in GloSea5 forecasts More variability is actually predictable Of course we hope to increase this in future forecast systems. Li et al 2016
Improved tropical teleconnections Yangtze is part of a banded structure of tropical rainfall Extends down to the deep tropics Poleward, moisture bearing winds occur to the south Li et al 2016
ENSO is a big part of this El Nino El Nino Xie et al 2009 Several mechanisms: IO memory Seasonal – ENSO interaction Stuecker et al 2015
Precipitation forecasts in 2016 From MARCH From MAY El Niño over by summer 2016 Very similar banded structure (Wet/Dry/Wet) Signals for wet in SE China
Real time forecast with China Met Administration (CMA) Philip Bett (Met Office) in collaboration with Chaofan Li (IAP) and Peiqun Zhang (CMA) Li et al, ERL, 2016 Useful regional average skill (r = 0.55) Real time service tested Wuhan flooding
Summer 2016 Yangtze Forecast Verification May-Jun-Jul precip was above-average and forecast was good
Summer 2016 Yangtze Forecast Verification Jun-Jul-Aug Precip was near-average Early forecasts were for wetter-than- average but the final forecast was close to average
Teleconnections of La Niña
This winter was more La Niña-like
Yangtze river summer precipitation following winter El Nino / La Nina Stephen Hardiman (Met Office) CSSP collaborative work with Chaofan Li (IAP) and Bo Lu (CMA) El Nino La Nina 0.24 mm/day -0.02 mm/day GPCP 0.28 mm/day 0.07 mm/day GloSea5 Yangtze river basin experiences anomalously wet summer following winter El Nino No signal in Yangtze summer precipitation following winter La Nina (response is NOT LINEAR) Response following both El Nino and La Nina is well simulated by GloSea5
El Nino La Nina Following winter El Nino, a strong anti-cyclone in the NWP drives northward, moisture bearing winds over the SCS. There is no signal in MSLP or meridional wind following a winter La Nina. Understanding why is the subject of further work (and may shed light on which mechanisms are important)!
ENSO in the previous winter has large influence on Yangtze summer rainfall What are the mechanisms?? Xie et al 2009 Stuecker et al 2015 Explain all mechanisms whilst on this slide... : Positive feedback from NE trade winds, getting stronger on Eastward flank (cold SST) and weaker on Westward flank (warm SST) Indian Ocean warms following El Nino, and keeps memory of warm SSTs longer. This produces Kelvin waves which anchor the AAC. Thus see a double peak in warm T in SCS (in winter and summer but not so much in spring) Alternatively, El Nino driven variability in NWP interacts non-linearly (i.e. constructive interference) with annual cycle of fields in that region – see peaks in the power spectra of the NWPAC (AAC) streamfunction at frequencies of 1-fe and 1+fe. i.e. this interaction acts to maintain the AAC. All the above seems to be about feedbacks between waves, circulation, and winds... Several possible mechanisms: Seasonal – ENSO interaction Local interactions in NWP IO memory of warm SSTs All lead to NWPAC, and thus poleward winds in SCS in JJA... May help to maintain NWP-AC Xie et al 2016
Evaluating the Risk of Extremes
South east England flooding www.theguardian.co.uk In south east England January 2014 saw the greatest monthly rainfall total on record Could it have been even worse?
Monthly rainfall totals In a given winter, there is an 8% risk of a month wetter than has been previously observed in south east England Thompson et al. submitted to Nature Communications
Hot summers in the Yangtze basin Vikki Thompson (Met Office) CSSP collaborative work with Hongli Ren and Bo Lu (CMA) Health impacts, with greater effect in urban areas Increases consumption of electricity and water Can lead to forest fires and crop losses Heat related mortalities in China have increased in recent decades What is the risk of a hotter summer month than has been seen?
Temperature data 40x more data available from model than observations Observations from WATCH Forcing ERA-interim Data (WFDEI): - Weedon et al. 2014 - Reanalysis data bias corrected with CRU TS3.1 - July and August monthly air temperatures - 35 x 2 = 70 months (years x months) Using Met Office decadal prediction system: - 40 realisations - 35 x 40 x 2 = 2800 months (start dates x realisations x months) 40x more data available from model than observations
Model fidelity Model and observations distributions appear similar
Model fidelity We resample the model 1000x and compare to the observations Model and observations distributions appear similar All measures must agree to 95% level
Only these regions pass the model fidelity tests for all measures Choosing the region Pass Fail Only these regions pass the model fidelity tests for all measures
Yangtze Climatology Monthly air temperature, black = observations, red = model
Yangtze Climatology Only July or August show record high temperatures Monthly air temperature, black = observations, red = model
Monthly temperatures Monthly air temperature for July and August black = observations, red = model
83 unprecedented extremes in the model Monthly temperatures 83 unprecedented extremes in the model In a given summer, there is a 6% risk of a hotter month than has yet been observed Monthly air temperature for July and August black = observations, red = model
Risk of an extreme in China 6% chance of a record each summer 1% chance each summer of July or August being 0.8 °C warmer than the current record
Monthly temperatures Top 10 August events Monthly air temperature, black = observations, red = model
2m air temperature standardised anomaly for top 10 August months Global temperatures 2m air temperature standardised anomaly for top 10 August months
2m air temperature standardised anomaly for top 10 August months Global temperatures 2m air temperature standardised anomaly for top 10 August months Hot Yangtze – more than two standard deviations above normal
Standardised precipitation anomaly for top 10 August months
Standardised precipitation anomaly for top 10 August months High rainfall over India,
Standardised precipitation anomaly for top 10 August months High rainfall over India, low rainfall over Yangtze basin
Conclusions The Met Office produces routine long-range forecasts using GloSea5 prediction system Outputs are available from the Met Office website and the Lead Centre for Long-Range Forecast Multi-Model Ensemble New developments are improving skill estimation and bias correction Research on various topics for European and global predictability is bearing fruit A number of topics on East Asian climate are being jointly tackled with collaborators at CMA and IAP under the UK-China Climate Science for Services Partnership (CSSP)
Thank You! Any Questions?
Skill for winter NAO is robust Winter (DJF) NAO r=0.53 GloSea5 skill level holds for: A longer (35 year) time series – using decadal version Similar skill for 1st and 2nd half of the period (Corr = 0.52 / 0.57) New model physics Decadal system DePreSys3; November starts; hindcast = 1980- 2014 Skill found with GloSea is robust – replicated over a longer period and with different model version Good news for NH extratropics! Nick Dunstone
Skill depends on ensemble size Forecast skill increases with ensemble size Occurs in other regional predictions Rate of growth depends on signal to noise ratio Li et al 2016, Scaife et al 2014