Convective-Scale Data Assimilation Dale Barker WWOSC, Montreal, Canada, 18 August 2014 Acknowledge: B. Macpherson, S. Ballard, G. Dow, R. Marriott, etc. © Crown Copyright 2014 Source: Met Office
Why Convective-Scale Data Assimilation? Flexibility/need to introduce observations more frequently (e.g. sub-hourly) than global DA. Simple downscaling -> spin-up problems Improved fit to observations through more accurate background forecast. Reduced obs representivity error. Additional, high-res ‘novel’ observations available (e.g. radar reflectivity)
Challenges For Convective-Scale DA Rapid cycling/output for fast processes Complex/Inhomogeneous observations. Imperfect high-res NWP model – model error. Limited predictability - probabilistic NWP/DA. Flow-dependent, non-Gaussian forecast errors. Careful treatment of multiple scales. Need to add value to global DA/NWP. Nonlinear processes (convection, microphysics) Earth System model coupling (e.g. land DA)
Novel Observations for Conv.-Scale DA http://wow.metoffice.gov.uk SEVERI Cloud Top Height: Radar (reflectivity, refractivity) TAMDAR: T, u/v, RH, icing, turbulence: Aerosol/visibility: Water Vapour Lidar: PV Cell (radiation):
TOMACS (Tokyo Metropolitan Area Convection Study) 2011 – 2013 Field Campaign Observations
UKV Observations (Day) SEVIRI SONDES SURFACE Anticipate beneficial impact from *4 increase in satellite data volumes when resolution is doubled
UKV Observations (Night) SEVIRI SONDES SURFACE Anticipate beneficial impact from *4 increase in satellite data volumes when resolution is doubled Need alternative sources of data at night: UAVs, night sonde launches, WV lidar?
Convective-Scale Predictability: EPS+ EDA (e.g Tong and Xue 2005) EnKF system based on ARPS OSSE experiments (using simulated data) for supercell storm dx = 2km, dz=500m Using 100 members Assimilate Vr and/or Z data Include 5 water/ice species
Single Reflectivity Observation: Multivariate Analysis Increments Shading : Full Fields Line Contours : Error Correlations Tong and Xue, 2005
2014: Status Of Real Data EnKF DA and Forecasting of Convective Storms Ming Xue (OU) 2014 EnKF workshop: ‘Obtaining good storm forecasts from relatively good EnKF analyses had been challenging for real cases; Forecast errors grew too quickly – deteriorates in 10-20 min; A number of later studies had only showed probabilistic forecasts that don’t have direct comparisons with observations; Model and environmental errors were believed to be key causes;’
Model Error in Tropical Convective-Scale NWP Stu Webster TOA OLR Spin-Up Issues (left) imply need for DA BUT large model biases (e.g. precip, above) need dealing with Current priority: Improve NWP model! Model Error is the elephant in the room for many CS-scale DA applications!
Met Office Main NWP Models (2014) UKV and MOGREPS-UK 1.5km 70L (40km model top) 3DVAR (3 hourly) 36hr forecast 8 times per day 12-member EPS - 2.2km 4x/day 36h UKV Radar MOGREPS-UK Global and MOGREPS-G 17km ENDGAME 70L (80km model top) Hybrid 4DVAR – 40km 66hr forecast twice/day 144hr forecast twice/day 44-member EPS - 33km 4x/day 168hr Probability Of Heavy Rain Red = Implemented in July 2014
Value Of Convective-Scale NWP Percentage benefit wrt UK Index* * UK Index = Forecast skill for surface weather: surface u/T, cloud fraction/amount, precip, visibility Global NWP improvements included in baseline above (~1-2%/yr). So 10% benefit of UKV represents > 5-10yrs lead over global model.
Value Of Convective-Scale DA NWP Range (UK Index): T+6 to T+36 DA = Cycling Convective-Scale DA, DS = Downscaler (Global DA) NoDA = No DA (forcing through LBCs) CS-scale DA significantly better than downscaler (DS) Complication: DA benefit includes cycling of prognostic aerosol in model.
Value Of Convective-Scale DA Nowcasting Range: T+6 to T+12 DA = Cycling Convective-Scale DA, DS = Downscaler (Global DA) NoDA = No DA (forcing through LBCs) Larger benefit of DA for nowcasting range (reduced impact of LBCs/model error)
Observations Assimilated Into UKV Model (updated September 2013) Anticipate beneficial impact from *4 increase in satellite data volumes when resolution is doubled * * Subset of data assimilated only in UK model
UK Observation Network denial experiments (Autumn period) Surface +2.9% Upper Air (excluding aircraft) +2.1% Aircraft +2.0% Radar Satellite +1.7% “Extra” (e.g. cloud) +0.5% Which observation types are most useful at convective-scale? These are some observation denial experiments in the 4km model (now retired). In this period, all major ob types are beneficial, with surface obs leading the way. Smaller benefit of extra mesoscale obs probably due to problem with MOPS cloud (since tackled by other developments with GeoCloud & surface cloud obs) Conclude: All ob types adding benefit, mainly from ‘standard obs’ in high-res DA.
Flow-Dependence Via An Adaptive Mesh Transform (Piccolo & Cullen, 2011: Q. J. R. Met. Soc., 137, 631-640) Computational mesh Nominal physical mesh Analysis Increment for single q ob above Sc band Motivation: lntroduce flow-dependence analysis response near strong temperature inversions in presence of stratocumulus clouds (hybrid 3DVAR problematic– too few members, large-scale perts). Static adaptive mesh methods concentrate grid points where there is a rapid variation of the atmospheric field. Transformation from the physical grid to the computational grid is guided by a monitor function: Grid transformation introduced within VAR control variable transform:
Iterative Calculation of Monitor Function UK4 domain: 3 Jan 2011 00z M (background-state - 3h forecast) M (After 10 iteration 3D-Var) M (After 2nd converged 3D-Var) Adaptive vertical grid provides a small positive impact to the UK index: Period Vis Precip Cloud amount Cloud base Temp Wind Overall 23 Dec 2010 – 3 Jan 2011 -2.56% 5.48% -1.05% 3.03% 0.22% -0.04% +0.25% 10 Aug 2010 - 20 Aug 2010 12.20% 0.00% 4.17% 0.23% 0.10% +0.55%
4D Variational Data Assimilation (4D-Var) (new) (initial condition for NWP) (old forecast) Challenges: Scientific (e.g representation of processes, linearity, model error) Practical e.g. very limited wall-clock available for NWP-nowcasting 4DVAR.
Convective-Scale Linear Model Test accuracy of linear model: 3.0km UM linearisation Tests T+3 (10th Mar 2012 – 06Z) T+3 (12th Mar 2012 – 00Z) U Theta Density Aerosol Accuracy of convective-scale linear model VERY situation dependent. Hourly 4DVAR less susceptible than adjoint-bases FSO tool.
Nowcasting Demonstration Project (NDP) 1.5 km NWP-based nowcasting system Southern UK only (May 2012 – April 2013) Hourly cycling 4DVAR (UKV=3hourly 3DVAR) Sue Ballard and team at MetOffice@Reading
NWP-Nowcasting: Precipitation Skill Verification of hourly precipitation forecasts against radar Same validity time, Available at same time to forecasters NDP better than older UKV forecast at all ranges NDP better than STEPS extrapolation/merged nowcast from T+2 Sue Ballard July 2012 August 2012 Fraction Skill Score (Roberts and Lean) for 1.0mm/h/40km square Against Forecast Range Next stage: UK-wide implementation of hourly 4DVAR in 2015-2016. © Crown copyright Met Office
Conclusions Predictability is limited to a few hours (at most) at convective-scale. Despite this, convective-scale NWP has shown great promise in recent years with significant human/HPC resource being devoted to it. Convective-Scale Variational Data Assimilation is the workhorse for current operational schemes, and contributes significantly to benefit of CS-scale NWP, BUT… Convective-Scale Ensemble(-Variational?) Data Assimilation likely to become more common in future - needs a (very expensive) convective-scale ensemble. Observation network limitations and model error perhaps largest challenges at present….
Thanks. Any Questions?
Tropical Balance Studies Dynamics of tropics VERY different to mid-latitudes. Traditional 3/4DVAR control variables inappropriate. Michel et al (2012) documented multivariate statistical regression linking all model variables. Here use 1.5km SINGV WRF output to train the regression. SIGNIFICANT multivariate correlations found - not represented in current global/high-res DA. <Chi, Chi_balanced> <T, T_balanced> <RH, RH_balanced> <PS, PS_balanced>