Predicting rainfall using ensemble forecasts Nigel Roberts Met Office @ Reading
MOGREPS-UK Convection-permitting 2.2 km ensemble now running routinely Embedded within MOGREPS-R ensemble members (18 km) 36-hour forecasts 12 members Every 6 hours Downscaling 18km initial conditions No high-resolution initial perturbations or forecast error perturbations to start with
Olympics showcase Probability of heavy rain Must have a convectionpermitting model to do this
Convective-scale ensembles (e.g. MOGREPS-UK) Where to begin? Are 12 members enough? How do we use it? How do we know how good it is? How can it be improved? Science and development?
Split forecasts into three scales Short to medium range weather forecasting 1 year 12-km forecasts Predictable scales (large synoptic) no need for an ensemble Uncertain scales (mesoscale) ensemble needed Noise (individual showers) neighbourhood processing with ensemble
Small uncertainty at large scales = large uncertainty at small scales Low 5% error at 1000 km = 100% error at 50 km
LBC driven uncertainties COPS IOP 8b at 12 UTC on July 27th 2007 Kirsty Hanley et al (2011) QJ PVU. Solid: six strongest convection members. Dashed: six weakest convection members (dashed)
Convective-scale ensemble How many members do we need? How many times do you need to throw a die for every number to come up - six? E = 6/6 + 6/5 + 6/4 + 6/3 + 6/2 + 6/1 = ~ 15 times N E= N 1/ i i=1 If 5x5km flood-producing storm is equally likely anywhere within 50 km square (not unrealistic). How many members to give a non-zero probability everywhere? Answer ~ 520 members (assumes perfect model and tessellation) Or ~ 2350 members if tessellate on 2.5 km grid 1% chance We only have 12 members!!
Reinvigorate your ensemble by giving it more members Use Neighbourhood processing it works like magic What happens at a particular model grid square is equally likely to occur at nearby grid squares This can be used to produce probabilities of occurrence. If 7 grid squares within a 5x5 grid-square neighbourhood around a particular grid square have rain, the probability of rain at that central grid square is 7/25. A 9x9 neighbourhood x 12 members = 972 members This is the neighbourhood 1000 deluxe ensemble Not independent members justifiable for unpredictable scales
Constructing a probability forecast Insufficient ensemble size leaves gaps
Constructing a probability forecast Probability of rain in period around the time of interest
Flooding near Aberystwyth 9 th June 2012 MOGREPS-UK probabilities at points ~32 km neighbourhoood Very predictable
Flooding near Aberystwyth 9 th June 2012 MOGREPS-UK probabilities within 20 km ~32 km neighbourhood Very predictable
Adaptive neighbourhood: Effect of ensemble size and resolution change
Short Term Ensemble Prediction System Scale decomposed radar extrapolation + high res NWP precipitation forecast Injects noise with space-time pdf from radar and high resolution NWP Errors modelled: Radar observation errors Extrapolation velocities Lagrangian evolution of extrapolated radar NWP forecasts Control Courtesy of Clive Pierce (Met Office) STEPS ensemble of surface precipitation rate 0300 UTC 17 November 2010
What do we want from a good ensemble? When the weather forecast is iffy we want plausible alternatives to know what might happen instead When the weather forecast is less iffy we should get less radical plausible alternatives In other words we want a good skill-spread relationship This has traditionally been measured at the gridscale or observation points (just like deterministic forecast verification). Is that helpful?
Spatial skill-spread using the FSS 99 th percentile, hourly accumulations Less spread More skill noise Spread, skill Large scale error Produced by Seonaid Dey, provided courtesy of Giovanni Leoncini
Hybrid Data assimilation Courtesy of Neill Bowler (Met Office) Dec 2010 Standard 3D-Var Pure ensemble 3D-Var Standard 4D-Var 50/50 hybrid 3D-Var u response to a single u observation at centre of window Became operational in Summer 2011 June 2011 Verification vs. obs
What do we do about an imperfect model? (Poor representation, biases and incorrect error growth) 1. Ignore it (errors grow quicker at high resolution Lorenz 1969) 2. Detect and correct/reduce (Thorwald) (some errors are not correctable) 3. Post process away (e.g. statistical downscaling) (only some errors) 4. Represent in ensemble somehow (e.g. stochastic representation of energy upscale from unresolved processes or uncertainties in physical processes) ECMWF improvements in skill and spread Jury out on tweaking physical parameters as in climate models
Physics perturbations compared to uncertainty through boundaries Taken from Gebardht et al, Atmos. Res. 2011 2.8 km model CR = N(GP all ) / N(GP >=1) CR = Correspondence ratio Number of pixels all members agree have rain divided by pixel with rain in at least one member Lower CR = more spread Very low rainfall threshold (Consistent with Vie et al 2011, MWR)
Physics vs. boundaries Seonaid Dey and Giovanni Leoncini FSS for precipitation hourly accumulations FSS Values 0-1 1 = perfect match 0 = totally different Time after start Contours every 0.1, colours black at 0.0 to red at 1.0 Graupel / convection scheme / timestep had little effect at reliable scales gridlengths
Ensemble comparison of microphysics schemes Taken from Clark et al, Jan 2012 BAMS Composite frequencies of observed rainfall greater than 0.50-in. relative to grid-points forecasting rainfall greater than 0.50-in. at forecast hour 30 from SSEF members using (a) Thompson, (b) WSM6, (c) WDM6, and (d) Morrison microphysics parameterizations. The boldface dot in each panel marks the center of the composite domain and the location of the observation.
Further research Where do errors / uncertainties grow and why Ensemble data assimilation (at convective and larger scales) Applying perturbations to convective-scale forecasts Relationship between scales How to verify convection-permitting ensembles How to post process and present Understanding the output Direct coupling to hydrological models Domain size, ensemble size and resolution Blending ensembles and seamless prediction
That s it