Initial trials of convective-scale data assimilation with a cheaply tunable ensemble filter

Transcription

1 Initial trials of convective-scale data assimilation with a cheaply tunale ensemle filter Jonathan Flowerdew 7 th EnKF Workshop, 6 May 016

2 Overall strategy Exploring synergy etween ensemles and data assimilation at convective scales: Flow-dependent covariances, convective-scale alances, interaction with orography, cross-system relationships,... Can ensemle DA improve UK deterministic/ensemle forecasts? Long-term research, eyond immediate operational plans such as UK hourly 4DVar Explore issues such as localisation, inflation, filter type, radar/cloud oservations, lateral oundary handling, updating hydrometeors, Initial experiments use a serial ensemle filter Results will inform ultimate operational ensemle DA system (e that EnKF or EnVar)

3 Update frequency The gold standard would e to synchronise the model with individual convective cells This requires perturations around the actual cell locations, which may e unlikely at T+3/6h for a small ensemle But might it e possile at T+1h, 15m, 5m? More linear, more Gaussian Many nudges within convective life-cycle, whilst preserving large scales Reduces time localisation issues Many EnKF studies use 5m updates May e limited y initialisation shock, or frequent small increments drawn from the same model may reduce such shocks 0 15m 1h 3/6h

4 Optimal static localisation Minimise RMS error, like the Kalman Filter For a single oservation: This also applies etween variales, etween model and oservations, etc Extra terms arise for dense oservations, eg: ) ( 1 f f a x y x x dz P J ) ( ) ( ) ( ) /(, N 1 1 o r Minimise T T f K K R H HP K K L ) ~ )( )( ~ ( ) J( Tr 1 ) ~ ( ~ ~ R H HP H P K T f T f f f P L P ~ Flowerdew (015), Tellus

5 Updating H(x f ) and a measure of analysis error Serial filters normally update x f for oservation j efore calculating H j+1 (x f ) Our separate OPS calculates all H(x f ) at the start using the original ackground state We can work around this y updating the oservation priors as additional elements of the state vector (Anderson, 003) a f Initial ackground error Diagnostic averages these errors As a onus, this naturally gives the innovation variance for each oservation after assimilating all prior oservations an independent measure of analysis error Final analysis error Oservation numer

6 First trial diagnostics June 014; sonde, surface and aircraft normalised innovations; 44+1 memers 6-hourly EnSRF Hourly EnSRF Hourly PertOs

7 Forecast performance Cycler worse than downscaler (loss of interior gloal DA?) 6h EnDA generally improves upon cycler 1h EnDA much etter than 6h Beats downscaler for temperature, not there yet for wind EnSRF and PertOs similar Also improves temperature ias Temperature Wind Pressure

8 Cheaply tuning the EnKF The RMS innovation after assimilation of prior oservations provides a way to cheaply tune many EnKF parameters To test the principle, run EnKF from archived input every 1.5d with all permutations of: Check y running full trial of suggested configuration The signal may e clear enough to uild tuning into cycle-ycycle EnKF MaxRadius (km) Equiv Gauss (km) Os/10k VertLocScale Gauss P factor Inf V \ H (data from Hourly PertOs ensemle mean)

9 Further trial results Tighter localisation improves performance, as suggested y tuner RTPP plus stochastic physics rings some further improvement Now often eating downscaler, despite limited oservation set Some improvement to categorical scores Temperature Wind Pressure Parameter Hourly EnSRF Tighter Loc RTPP Downscaler 0/6/1/18 UTC Surface visiility h Precip Accum Total Cloud Amount Cloud Base Height (3/8 cover) Equitale Threat Scores (higher is etter)

10 SEVIRI cloud-affected radiances Pete Weston Useful oservation type to test: Satellite assimilation normally ignores areas affected y cloud Cloud is a key forecast variale, tied to convectivescale features Proes dense oservations, awkward variale, non-trivial oservation operator Satellite Applications independently chose to test ensemle filter Natural synergy etween plans Previous CsEnDA system, plus: channel 5 (upper tropospheric humidity, red) channel 9 (cloud top, or surface if clear, cyan) SEVIRI localisation can differ from conventional os Increment cloud water and ice

11 SEVIRI trial results Strongly draws towards susequent oservations Incrementing cloud water/ice is slightly detrimental (need inter-variale localisation?) Adding just channel 5 is eneficial Adding channel 9 harms overall performance (vertical localisation needs to move with the covariances?) Tuner suggests SEVIRI horizontal localisation narrower than conventional oservations Tuner suggests vertical localisation roadening similar to typical Jacoian widths 111 km Narrow Suite ID Description Ctrl UKV Cloud Cloud ase ID index amount height mi-af844 Conventional os N/A N/A N/A N/A mi-ag759 Increment cloud water/ice af % mi-ah038 Only channel 5 ag % Broad mi-ag84 Channels 5 & 9 ag % mi-ah361 Broad localisation ag % km

12 Conclusions The serial filter is a promising technique for convectivescale data assimilation A complete CsEnDA suite has een developed, with flexile cycle length Hourly assimilation performs much etter than 6-hourly Both the deterministic and pertured oservation filters are worth considering Tighter horizontal and vertical localisation is eneficial SEVIRI channel 5 is eneficial, channel 9 more challenging The RMS innovation after assimilating prior oservations (IAPO) is a useful diagnostic and allows cheap tuning of parameters such as localisation radii It may also provide a way to calculate the relaxation factor without having to know the oservation error

13 Future work Further SEVIRI work: Further trials, tuning, diagnostics Vertical localisation moving up/down for channel 9? Apply tuner to oservation errors? Shorter (15/30 minute) cycles? Inter-variale localisation? Complete wider EnKF experiments: Ensemle verification Localisation, inflation/relaxation, LBCs,... Radar assimilation Diagnostics/case studies Comparison to 3/4DVAR, perhaps LETKF Proposed PhD project extending theoretical/idealised work on localisation, inflation/relaxation and serial/parallel filters Then decide what to uild operationally

14 Any questions? Thanks to: Neill Bowler, Gordon Inverarity, SA Cloud Analysis Review Group, Susanna Hagelin, Kelvyn Roertson, Gareth Dow, Anne McCae, Jorge Bornemann, Adam Maycock, David Davies, Ro Darvell,...