Introduction. Tilly Driesenaar

Transcription

1 Tilly Driesenaar Introduction Tilly Driesenaar The first of September this year the HIRLAM program celebrated his 2 th anniversary. This occasion nicely coincided with a visit of the management group to DMI, where it all started September 1st 198. For the occasion a picture was taken of (part of) the management group together with Bent Hansen Sass who also was there at the start 2 years ago. This Newsletter contains a number of scientific contributions and a few reports of visits to member institutes. It s good to see that the focus is shifting towards Harmonie in the contributions. First we see an article by Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrae who compared the 3-dimensional variational data assimilation schemes in HARMONIE and HIRLAM. The next contribution, also by Magnus Lindskog deals with ALADIN 3D-VAR utilizing a wide extension zone. Then Wim de Rooy tells about the experiences with Harmonie at KNMI. A revised method to determine the hybrid coordinate in HIRLAM is presented by Per Undén and Huseyin Toros, Gertie Geertsema and Gerard Cats contribute with an evaluation of the precipitation forecasts of HIRLAM and Harmonie for a flash flood event that occurred in Istanbul in September 29. Then Newsletter is concluded by reports of the visits of the management group to AEMET, FMI and MetÉireann. 1

2 2

3 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ A comparison of Harmonie and Hirlam 3-dimensional variational data assimilation 29 September 29 Magnus Lindskog 1, Sigurdur Thorsteinsson 2 and Ulf Andræ 1 1 Swedish Meteorological and Hydrological Institute, Norrköping, Sweden 2 Icelandic Meteorological Office, Reykjavik, Iceland 1. Introduction The ALADIN data assimilation system has been incorporated into to the HARMONIE mini-sms environment. The main components of HARMONIE mini-sms are the variational assimilation, the surface assimilation (CANARI) and the forecast model. The system also includes observation processing, postprocessing, monitoring and verification components. As a first investigation of the functionality and performance of ALADIN 3D-Var within the HARMONIE system it has been set-up over an area roughly coinciding with the HIRLAM RCR area. Figure 1 shows the horizontal extent of the HARMONIE model domain (extension zone included) as compared to the HIRLAM RCR model domain. The horizontal resolution is roughly 0.1 degrees and the same 60 vertical levels are used for both HARMONIE and HIRLAM. HARMONIE RCR uses polar stereographic projection while HIRLAM RCR uses rotated latitude-longitude projection. The HARMONIE experiment is using version 33h1 with the ALADIN physics in the forecast model. The HIRLAM version used is Both HARMONIE and HIRLAM use 3D-Var. The experiments were run on the ECMWF hpce platform. For both systems, only in-situ measurements were assimilated. For the ALADIN 3D-Var an extension zone of 11 grid-points ( ~170kms ) is used and no observations closer to lateral boundaries than 1 kms are used, for HIRLAM the corresponding extension zone is around 9 kms. 2. Background error statistics At present, before being able to carry out HARMONIE data assimilation on a new model domain, background error statistics need to be generated. Background error statistics for HARMONIE have been generated from differences of 4 parallel one month HARMONIE runs in forecast mode only. Figure 1. Horizontal extent of the HIRLAM RCR area (red) and HARMONIE area (blue). 3

4 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ Figure 2. Vorticity mean vertical correlations as function of vertical level (level 1 is the uppermost and level 60 is the lowest model level). The 4 different runs are started once a day from 6h forecasts based on analyses of different members of an ECMWF data assimilation experiment, with perturbed observations and model physics. Lateral boundary conditions were taken from the corresponding ECMWF forecasts. A total of h forecast differences were generated and used in the generation of background statistics. Like in HIRLAM, a statistical balance formulation is used (Berre, 20). However, unlike in HIRLAM, in HARMONIE a constant Coriolis parameter is used in the derivation of the statistics. Background error standard deviations are scaled with a parameter 1.9 to obtain roughly the same order of magnitude as HIRLAM background error standard deviations when using a scaling factor of 0.9 (the value that should be used when large scale mixing is not applied) which is the same value applied in synoptic scale comprehensive impact studies. As an example of derived HARMONIE background error correlations, the vorticity mean vertical correlations are shown in Figure 2, as function of vertical level. The vertical correlation length scales are smaller at higher altitudes (lower number of model level). The results are in accordance with previous studies presented in the literature (i.e. Fig. in Berre, 20). Background error standard deviations for the control variables vorticity and, unbalanced temperature are shown for HARMONIE and HIRLAM in Figure 3. The HIRLAM structure functions are based on the NMC-method and are constructed from statistics of differences of +36 and +12 h forecasts, valid at the same time. For vorticity, the background error standard deviations are decreasing towards the top and the bottom of the model, both for HARMONIE and HIRLAM. Some small differences exist for the structure of vorticity background error standard deviations. The standard deviations of the HARMONIE unbalanced temperature decrease towards the top of the atmosphere, both for HARMONIE and HIRLAM. For HARMONIE there is in addition a decrease of unbalanced temperature background error standard deviation towards the surface. The reason is probably that the perturbed surface temperatures from the ECMWF data assimilation experiment are not used (for technical reasons). For vorticity the HARMONIE background error standard deviations are slightly smaller than the HIRLAM ones. For unbalanced temperature the HARMONIE background error standard deviations are slightly larger than the HIRLAM ones. The HARMONIE structure functions are further investigated and compared with the HIRLAM ones through single observation impact experiments. One single temperature observation at hpa and 3 K warmer than the corresponding background value was placed in the Atlantic, outside Great Britain. The horizontal influence of the single temperature observation on the temperature and wind field at model level 23, close to hpa, is shown in Figure 4. For both HARMONIE and HIRLAM 4

5 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ Figure 3. HARMONIE (dashed red line) and HIRLAM (full blue line) vertical profiles of scaled background error standard deviations. For HARMONIE the scaling factor is 1.9 and for HIRLAM the scaling factor is 0.9. Vorticity (left) and unbalanced temperature (right). Figure 4. Horizontal impact on temperature ( unit:k times 10) and wind at model level 23, due to one single temperature observation at hpa. Left HARMONIE and right HIRLAM. the multivariate structure functions induce a anti-cyclonic circulation around the observation. The significantly shorter horizontal correlation length scales for HARMONIE are due to the ensemble based structure functions as compared to the NMC-based for HIRLAM. These shorter correlation length scales are regarded as something good, since the NMC method is known to overestimate the correlation length scales. The vertical impact on temperature due to the single temperature observation is demonstrated in Figure 4, which show a cross-section in a area close to the observation. It can be noted that also vertical correlation length scales are shorter for HARMONIE, due to the ensemble based structure functions.

6 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ Figure. Vertical impact on temperature (unit:k times 10) due to one single temperature observation at hpa. Positive increments with full red lines and negative increments with dashed blue lines. Left HARMONIE and right HIRLAM. 3. Diagnosis of one single data assimilation cycle 3.1 Observation usage Table 1 provides detailed information of observation usage for HARMONIE and HIRLAM for the cycle UTC. The observation usage is quite similar for most types of observations. However, some differences exist and these can be explained. Much less radiosonde humidity data are used in HARMONIE. The reason is that no humidity data from radiosondes above 3 hpa are used in HARMONIE. The quality of the radiosonde humidity observations at these high elevations are considered to be too poor, which is probably true. HARMONIE utilizes much more PILOT u/v because American windprofilers were used by HARMONIE but not by HIRLAM. None of the systems used EUROPEAN windprofilers. More AIREPS were used in HARMONIE. The reason is probably the small thinning distance set in Observation type Harmonie Hirlam SYNOP z SHIP z DRIBU z TEM T TEMP u/v TEMP q PILOT u/v AIREP T AIREP u/v Table 1. Observation usage in one single data assimilation cycle UTC. 6

7 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ Figure 6. HARMONIE (left) and HIRLAM (right) SHIP surface pressure observation usage and rejections for the assimilation cycle UTC. Blue dots means active, red dots indicate rejected and black dots passive. namelist for HARMONIE (2 km). More SHIPs and DRIBUs were used in HIRLAM than HARMONIE. The reason is that HARMONIE uses an IFS thinning distance of 1 degree while HIRLAM has half a grid distance. Figure 6 shows horizontal maps of SHIP surface pressure observation usage and rejections for HARMONIE and HIRLAM. 3.2 Observation fit statistics Very small adjustments have been done to observation error statistics of the HARMONIE to be close to the HIRLAM (and ECMWF) statistics. As a first step also the background error statistics of HARMONIE were adjusted (using a single scaling factor, see previous section) to be as close to the HIRLAM statistics as possible. Observation fit statistics for temperature and u wind component observations from radiosondes are shown in Figures 7 and 8, for UTC. The same ECMWF forecast was used as background state for both HARMONIE and HIRLAM. It can be seen in Figures 7 and 8 that the observations have a similar influence in HARMONIE and HIRLAM data assimilation, as indicated by the difference between rms of the the ob-fg and ob-an departure. The large bias and rms values in the uppermost layers can be explained by a small sample size. 3.3 Analysis increments Figure 9 shows HARMONIE and HIRLAM Figure 7. HARMONIE (upper) and HIRLAM (lower) radiosonde temperature observation fit statistics for the assimilation cycle UTC. Bias and rms of observation minus background (blue dotted line) and observation minus analysis (red full line).unit: K. 7

8 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ temperature analysis increments at model level 23 (close to hpa) for UTC. The same ECMWF forecast was used as background state for both systems. Many features of the analysis increments of the two systems are rather similar to each other, and do roughly the same job when updating the background state with information from observations. Some differences are found. These are probably mainly due to differences in background error statistics and quality control. Differences are also found close to lateral boundaries. These are not visible in Figure 9 due to the choice of having rather few isolines for a clear comparison of the main features. However for HARMONIE there are spurious relatively small amplitude increments propagating from one side of the domain to the other. These are due to the bi-periodicity of the spectral representation in combination with the relatively narrow extension zone of 11 grid-points. Another related issue is of course the horizontal correlation length scales. These spurious increments are not present in the HIRLAM analysis increments, due to the larger extension zone. 4. One week parallel data assimilation and forecast experiment A one week parallel forecast and assimilation experiment has been carried out for the period Figure 8. HARMONIE (upper) and HIRLAM (lower) radiosonde u-wind component observation fit statistics for the assimilation cycle UTC. Bias and rms of observation minus background (blue dotted line) and observation minus analysis (red full line). Unit: m/s. Figure 9. Temperature analys increments at model level 23 (close to hpa) for UTC. Unit: Kelvin times 10. HARMONIE (left) and HIRLAM (righy). 8

9 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ Figure 10. HARMONIE (red) and HIRLAM (blue) bias and rms scores for verification of forecasts against observations, as function of forecast length.upper left: 2 hpa temperature (unit:k). Upper right: 3 hpa wind speed (unit: m/s). Lower left: 7 hpa temperature (unit:k). Lower right 7 hpa wind speed (unit:m/s) Scores for verification of forecasts against observations (Figures 10 and 11) indicate that the HARMONIE forecasts are slightly better up to 6-7 hpa. Above 6 hpa the HIRLAM forecasts have somewhat better quality. One should keep in mind that the period of the experiment is far to short for deriving statistically significant differences. It should also be kept in mind that the forecast models in HARMONIE and HIRLAM are different.. Conclusions The study aimed at evaluating the functionality of the components of the HARMONIE data assimilation system. Observation handling of conventional types of observations is very similar in the two systems. Some differences are found and can be explained. The rejection of humidity observations above 3 hpa in HARMONIE is considered appropriate, considering the present quality of these observations. The HIRLAM thinning distances are specified in number of gridpoints and thus resolution dependent. It is probably better to have the thinning distances specified in terms of absolute distances, like in HARMONIE. The main purpose of thinning is to remove observation error correlations, that can presently not be handled by the data assimilation system, and these should be model resolution 9

10 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ Figure 11. HARMONIE (red) and HIRLAM (blue) bias and rms scores for verification of forecasts against observations, as function of vertical level.left: temperature (unit:k). Right: wind speed (unit: m/s). The scores are summarized over +12, +24, +36 and +48 h forecasts. independent. In addition the thinning reduces the number of observations to an amount that can be handled, and the influence of a particular observation type within the data assimilation. The absolute values of the thinning distances used in HARMONIE presently could therefore be investigated in more detail (like the ones in HIRLAM). The next step for HARMONIE should be to introduce various types of remote sensing and local observations. The background error statistics are in general rather similar in HARMONIE and HIRLAM, and are in both cases based on a statistical balance formulation. A limitation of the HARMONIE structure functions is the use of a constant Coriolis parameter, while this not the case in HIRLAM. Using a constant Coriolis parameter is a questionable approximation for large model domains. The constant Coriolis parameter approximation is however not used in HARMONIE if the non-linear balance equation (and omega equation) is activated in the background error statistics. However, in the HARMONIE version of the non-linear balance equation model levels are assumed to be approximately on levels of constant pressure, which is not always true. The effect of this assumption should be further investigated and if possible improved. It should also be mentioned that it would be desirable to as a first step be able to interpolate structure functions from one domain to another, like in HIRLAM. The main features of the assimilation increments look very similar in HIRLAM and HARMONIE. However, spurious HARMONIE assimilation increments close to lateral boundaries are found. These 10

11 Magnus Lindskog, Sigurdur Thorsteinsson and Ulf Andrœ are due to the bi-periodocity of the spectral formulation and due to the relatively narrow extension zone. Related is the horizontal length scale of the structure functions. This is regarded as a more serious problem that needs to be taken care of. When increasing the model resolution and having a small domains this problem will be even more pronounced. It can be avoided by increasing the number of grid-points in the extension zone. So far most (if not all) HARMONIE experiments have been carried out with a 11 grid-point extension zone. A problem is that forecast model calculations are carried out in the extension zone. This will result in very high (unnecessary) computational costs if not taken care of when increasing the extent of the extension zone. A related issue (but not HARMONIE specific) is that more refined methods for calculating structure functions (ensemble data assimilation) should be applied for high resolution domains. At present the wrap-around problem can be reduced by not using observations closer to lateral boundaries than 20 kms (instead of the presently used 1 kms). This is however not regarded as a satisfactory long term solution. The one week parallel experiment indicated that in terms of scores for verification of forecast against observations, the HARMONIE and HIRLAM systems used were of roughly similar quality. HARMONIE performed slightly better at lower vertical levels while HIRLAM performed better at higher vertical levels. It s difficult to attribute the differences to any part special part of the system and it should be kept in mind that these scores are based on a one week period only. References Berre, L., 20. Estimation of synoptic and meso scale forecast error covariances in a limited area model. Mon. Wea. Rev., 128,

12 Magnus Lindskog Aladin 3d-var with a wide extenzion zone Magnus Lindskog Swedish Meteorological and Hydrological Institute 4 February Introduction The 3-dimensional variational (3D-Var) data assimilation of the ALADIN forecasting system is utilizing a bi-fourier representation of model fields. To obtain periodically varying fields over the limited area model domain an extension zone is applied. The width of the extension zone should allow horizontal background error correlations to approach zero across the extension zone. The reason for using a wide extension zone is to avoid spurious wrap-around of data assimilation increments from one side of the model domain to the other. The typical horizontal length scales of the background error statistics are dependent on type of variable and vertical level (larger lengths scales at higher altitudes) but are at least around 6-7 kms. Usually an extension zone of 11 gridpoints is used for ALADIN. Considering a horizontal resolution of to 10 kms this will result in an extension zone width of to 110 kms. This is far too narrow to avoid spurious wrap-around of data assimilation increments. In principle the number of gridpoints in the extension zone could be increased in ALADIN. The problem is that in the forecast step (and in 4D-Var), following the data assimilation, model calculations are carried out also in the extension zone. Extending the number of gridpoints in the extension zone will thus significantly increase the computational cost. There is however on-going work to get rid of the model calculations in the extension zone, since these are un-necessary. In the present study the ALADIN forecasting system is set-up over a Scandinavian domain, utililizing a wide extension zone. The set-up is referred to as EXT. The functionality and performance of EXT is evaluated and compared with the one of a reference set-up, CRL, utilizing an 11 gridpoint extension zone. The specifications of EXT and CRL differ only in the width of the extension zone, the inner areas of EXT and CRLare very similar. All other settings are exactly the same in EXT and CRL. 2 Preparations and Experimental design 2.1 General set-up The ALADIN forecasting system, including 3D-Var, has been set-up over an Scandinavian domain. The horizontal resolution is. kms and 60 vertical levels are used. The forecast model is run in hydrostatic mode applying ALARO physical parameterization. In EXT a 120 gridpoints wide extension zone (660 kms) has been applied. For comparison purposes, also an alternative set-up, CRL, with an 11 gridpoints wide extension zone ( kms), has been set-up. The set-ups differ only in the width of the extension zone. The inner areas of EXT and CRL are very similar and the rest of the settings are identical, except for the background error statistics, that slightly differs. The CRL domain is the one used in pre-operational studies at the Swedish Meteorological and Hydrological Institute (SMHI). The set-ups of the domain with the wider extension zone (EXT) and the run with 12

13 Magnus Lindskog Figure 1: Scandinavian domain for EXT (left, red: inner area and black: extended area) and CRL (right, blue: inner area and black: extended area). the narrower extension zone (CRL) are illustrated in Figure 1. In Figure 2 it is demonstrated that the inner areas of EXT and CRL are very similar. There are however, some slight differences in the inner domain. In particular the location of the Northern and the Southern lateral boundaries differ, but very slightly. For the data assimilation, conventional types of observations are used. No observations closer to the lateral boundaries of the inner area than 1 kms are assimilated. This is a common procedure to allivate to some extent the spurious wraparound of assimilation increments, compensating for a too narrow extension zone. The negative effect is that no information from observations closer to the lateral boundaries of the inner domain than 1 kms are taken into account. Furthermore for both EXT and CRL the somewhat empirical scaling factor used for background error statistics (REDNMC) was set to 1. Figure 2: Inner areas of Scandinavian domain for settings EXT (red) and CRL (blue). 13

14 Magnus Lindskog 2.2 Background error statistics Since the extended domains are different, pre-calculated background error statistics are required both for EXT and CRL. For each of the two domains, the statistics are calculated from an ensemble of 6 h forecasts. These are launched from 6 h forecasts of an ECMWF ensemble data assimilation experiment with perturbed observations and model physics parameterizations. The ECMWF ensemble generated forecasts are also used as lateral boundary conditions for the ensemble of ALADIN forecasts. There are 4 ensemble members run for the period of to Statistics are calculated based on the differences between the four members during the period to Forecasts are only launched from UTC. The derived background error standard deviation profiles for vorticity and unbalanced divergence are presented in Figure 3. It can be seen that the shape of the vertical profiles is very similar, but the magnitude of the background error standard deviations are smaller for EXT than for CRL. The derived background error standard deviations for unbalanced temperature and unbalanced humidity are presented in Figure 4. For unbalanced humidity the shapes of the vertical profiles are again very similar, but with smaller values for EXT. For unbalanced temperatures the shapes of the vertical profiles are also rather similar, but with slightly smaller EXT values at higher vertical levels and slightly lower EXT values at lower vertical levels, as compared to CRL. For surface pressure background error standard deviations (not shown) the EXT values are slightly larger than the CRL values (roughly 1 percent). The differences in background error standard deviations are due to the fact that the statistics is not based only on forecast differences in the inner area. It is also based on forecasts differences in the extension zone, which is much wider for EXT than for CRL. The HIRLAM approach is to set the forecast differences to zero in the extension zone. In our case differences of extrapolated fields (extension zone) contribute with a much larger part to the total difference field in EXT than in CRL. It seems thus from Figures 3 and 4 that the difference fields in the extension zone of the EXT domain are smaller for vorticity and divergence, but larger for surface pressure. The latter may be caused by the fact that the biperiodication of the spatially inhomogeneous surface pressure field over a large extension zone causes large differences and thus large background error standard deviations. A comparison of covariances and correlations of EXT and CRL reveals that the main features are Vertical level 30 Vertical level Vorticity x Unbalanced divergence x 10 Figure 3: Vertical profile of background error standard deviations for vorticity (left, unit: s 1 ) and unbalanced divergence (right, unit: s 1 ). Red full curves are for EXT and dashed blue curves are for CRL. 14

15 Magnus Lindskog Vertical level 30 Vertical level Unbalanced temperature Unbalanced humidity x 10 4 Figure 4: Vertical profile of background error standard deviations for unbalanced temperature (left, unit: K) and humidity (right, unit: kgkg 1 ). Red full curves are for EXT and dashed blue curves are for CRL. 1 T du covs ext x T du covs crl x T vertical level T vertical level divu vertical level divu vertical level Figure : Mean vertical crosscovariance matrix between temperature and unbalanced divergence (units: Ks 1). Left is for EXT and right is for CRL. very similar. As an example the cross correlations between temperature and unbalanced divergence are shown in Figure. The results agree very well with the ones presented in Figure 11 b in Berre (20). 2.3 Parallel experiment To evaluate the functionality of EXT and the differences as compared to CRL a 18 day parallel data assimilation and forecast experiment has been carried out. The experimental period is from UTC to UTC. During this period EXT and CRL are run within a 6 h data assimilation cycle and at and at 12 UTC 12 h forecasts are launched. The reason for limiting the forecast range to 12 h is the high computational cost of EXT forecasts. ECMWF operational forecasts are used as lateral boundary conditions and conventional types of observations are assimilated. The parallel experiment has been carried out at the ECMWF c1a platform. 1

16 Magnus Lindskog 3 Results 3.1 Study of single assimilation cycle For the very first data assimilation cycle UTC EXT and CRL utilize the same background state, which is a 6 h ECMWF forecast, interpolated to the model domains. The surface pressure analysis increments for EXT and CRL for the first data assimilation cycle are shown in Figure 6. As expected, in most parts of the domain the increments are very similar between EXT and CRL. However in data sparse areas near the Northern lateral boundary and the Northern part of the Western lateral boundary some signifcant differences occur. The differences extend quite a bit into the domain. By inspecting the increments at the Southern and Southern part of the Eastern lateral boundary it is obvious that the differences are mainly related to the wrap-around of analysis increments in CRL. 3.2 Mean analysis increments In Figure 7 the surface pressure mean analysis increments for EXT and CRL, accumulated over the period to are shown. Again, over the inner part of the domain and in data dense areas the mean assimilation increments are very similar. However, close to the data sparse Western lateral boundary the increments are significantly different. For EXT, negative mean assimilation increments are found in the lower left corner of the domain and close to the the Island of Jan Mayen, at the upper left corner of the domain. The latter are probably caused by systematic representativity errors of the surface pressure observations at the station of Jan Mayen. For CRL, however, due to wrap-around effects the mean assimilation increments are positive in these two areas, at the lower left and upper right corner of the domain. For CRL the systematic positive surface pressure increments (and systematically negative temperature increments, as is shown below) in the lower right corner spread also to the lower left and upper left corner. For CRL the wrap-around effect dominates over the effect of the systematic representativity error of the observation at the John Figure 6: Surface pressure analysis increments at UTC for EXT (left) and CRL (right). Unit: hpa times

17 -10 HIRLAM Newsletter no 6, November 2010 Magnus Lindskog Figure 7: Surface pressure mean analysis increments at accumulated from UTC to UTC. Left is for EXT and right for CRL. Unit: hpa times 1. Mayen station. The lattter is therefore not visible in the CRL mean surface pressure assimilation increments. The temperature mean assimilation increments at model level 20 (slightly above hpa) are shown in Figure 8. Again the increments are accumulated over the period UTC to Figure 8: Mean analysis increments for temperature at model level 20 accumulated from UTC to UTC. Left is for EXT and right for CRL. Unit: mk times 1. 17

18 Magnus Lindskog stations Area: ALL Temperature Period: At, No cases RMSE 3dext RMSE 3de11 BIAS 3dext BIAS 3de11 CASES stations Area: ALL Wind speed Period: At, No cases RMSE 3dext RMSE 3de11 BIAS 3dext BIAS 3de11 CASES hpa hpa deg C m/s Figure 9: Time-averaged bias and rms of EXT (red) and CRL (blue) 12 h forecasts as function of vertical level. The scores are for temperature (upper, unit: K) and wind speed (lower, Unit: m/s) time-averaged over the period UTC to UTC. 18 UTC. Again the main features of the increments are very similar between EXT and CRL in the central part of the domain and close to the data dense Eastern and Southern parts, although some differences exist. Significant differences can be found close to the data sparse Northern and Western lateral boundaries. In particular it is obvious that the systematic negative CRL increments in the North-Western and South Western corner of the domain are due to spurious wraparound effects. They result from the systematically negative assimilation increments in the South-Eastern part of the domain. These spurious temperature mean analysis increments at the North-Western and South-Western corner can be found at most vertical levels. Furthermore temperature-surface pressure correlations are spreading these spurious increments also to surface pressure, which causes the differences in the surface pressure mean analysis increments in the North-Western part of the domains in Figure 3.3 Scores for verification against observations The forecasts of the two parallel runs have been verified against observations for the period Area: ALL 66 stations Temperature 7 hpa At, Window: 12h RMSE 3dext RMSE 3de11 BIAS 3dext BIAS 3de11 CASES Area: ALL 6 stations Wind speed 80 hpa At, Window: 12h RMSE 3dext RMSE 3de11 BIAS 3dext BIAS 3de11 CASES deg C 0.4 No cases m/s 1 0. No cases /02 12/02 14/02 16/02 18/02 20/02 Date 22/02 24/02 26/02 28/ / /02 12/02 14/02 16/02 18/02 20/02 Date 22/02 24/02 26/02 28/ /03 Figure 10: Temperature bias and RMS for 7 hpa 12 h forecasts for EXT (red) and CRL (blue) as function of data assimilation for the period to Unit: m/s. Figure 11: Wind speed bias and RMS for 80 hpa 12 h forecasts for EXT (red) and CRL (blue) as function of data assimilation for the period to Unit: m/s. 18

19 Magnus Lindskog Area: ALL 1064 stations Surface pressure At, Window: 6h RMSE 3dext RMSE 3de11 BIAS 3dext BIAS 3de11 CASES hpa No cases /02 12/02 14/02 16/02 18/02 20/02 Date 22/02 24/02 26/02 28/ /03 Figure 12: Surface pressure bias and RMS for EXT (red) and CRL (blue) as function of data assimilation for 12 h forecasts for the period to Unit:hPa UTC to UTC. They are presented in Figure 9 for temperature (upper) and wind speed (lower). In terms of root mean square error (rms) the temperature forecasts are better for EXT than for CRL below hpa. Above hpa the scores are neutral. On the other hand the wind speed forecasts are slightly worse for EXT, at 80 and 2 hpa. At other vertical levels the scores are neutral. In Figures 10 and 11 the verification scores are shown as function of data assimilation cycle, for 7 hpa temperature and 80 wind speed, respectively. It can be seen that the temperature scores are considerably better for EXT than for CRL. For wind speed scores however, during some period CRL performs better and during others EXT performs better. However, as shown in Figure 12, the surface pressure verification scores are systematically better for CRL than for EXT. Possibly this is relate to wrap-around in CRL compensating for the systematic representativity error at the Jan Mayen station. 4 Conclusions The ALADIN forecasting system with application of a wide extension zone has been set-up over a Scandinavian domain. The system is technically working and has stably run for an 18 day period, although computationally expensive. A comparison with a set-up with 11 grid-point extension zone reveals benefits of using a large extension zone. The spurious wrap-around of analysis increments is avoided. The effects of the wrap is primarily noticable close to data sparse lateral boundaries, but due to propagation and cycling there will be a smaller effect also in the inner part of the domain. An 18 day parallel forecast and assimilation experiment clearly shows that the width of the extension zone has an effect also on the forecast quality in the entire domain. So the effect of the width of the extension zone (including differences in background error statistics caused by it) are not only seen close to the lateral boundaries. Probably the impact is even larger for longer range forecasts than the 12 h forecasts used in this study. With longer forecasts even more of the spurious analysis increments mainly located close to the lateral boundaries reach the inner part of the domain. In particular the temperature and humidity (not shown) forecasts are improved when appling a wide extension zone. The surface pressure forecasts seem to get slightly worse for this period. Possibly this is related to wrap-around in CRL compensating for the systematic representativity error at the Jan Mayen station. This issue needs to be further investigated and understood in future studies. From this study it is obvious that a narrow extension zone causes serious problems, at least for some choices of model domain and observation usage. Probably the negative effects of the extension zone would be alleviated if choosing a domain with a more uniform observation 19

20 Magnus Lindskog density and/or include satellite observations. However, inhomogeneties will always exist for some times and areas. Due to the results obtained here it is recommended to continue the work with avoiding unnecessary forecast model calculations in the extension zone. Thereafter more extensive extended parallel experiments and further investigation of the impact of the width of the extension zone on forecast quality should be done. Acknowledgements Nils Gustafsson, Per Undén and Loïk Berre are greatly acknowledged for reading through and for commenting on the manuscript. References Berre, L., 20: Estimation of synoptic and meso scale forecast error covariances in a limited area model. Mon. Wea. Rev., 128,

21 Wim de Rooy et.al. Experiences with Harmonie at KNMI Wim de Rooy Cisco de Bruijn, Sander Tijm, Roel Neggers, Pier Siebesma, Jan Barkmeijer Introduction This report describes the experiences at KNMI with the Harmonie model as provided by continuous 1D verification in the KNMI parameterization Testbed 1) and 3D runs of several cases. Although the validation period is rather short and the results have to be investigated in further detail, some features are that noticeable that they are described here. For shortness we will describe only two cases here but the results are representative for our experiences so far. For the turbulence/convection (EDMF) scheme two different options are investigated: EDKF: the default scheme at Meteo France 2) 3), 4), ) EDMFm: a dual updraft EDMF scheme developed at KNMI Soon after experimenting with Harmonie it became clear that the default EDKF settings for the cloud scheme (i.e. CMF_CLOUD=DIRE and LOSIGMAS=TRUE) results in too low variances for the statistical cloud scheme and consequently too much all-or-nothing cloud fraction behavior. The default setting for EDMFm was (in version 3h1.2) CMF_CLOUD=STAT and LOSIGMAS=FALSE which invokes a Chaboureau & Bechtold 6) parameterization for the variance. Unfortunately this setting turned out to lead to opposite problems showing too much variance with consequently too large areas with clouds. Moreover, such a statistical cloud scheme does not provide a physical coupling with the turbulence and/or convection scheme. Considering the above we decided to modify the statistical cloud scheme in Harmonie. First of all we use the combination CMF_CLOUD=STAT and LOSIGMAS=TRUE which includes the physical coupling between the statistical cloud scheme and the turbulence as well as the convection scheme (following ideas of 7) ). Additionally we modified the cloud scheme as explained in the next section to come to a more realistic cloud cover fraction. Results of EDMFm with this modification, noted as EDMFm+ will be compared to results with the default EDKF options in the results section. Modifications to the cloud scheme The key parameter in every statistical cloud scheme is the variance of variable s. Variable s is the distance to the saturation curve, s q 1 - q sat (P, T), where q t is the total water specific humidity, q sat is the saturation total water specific humidity, P is the pressure, and T is the temperature. The variance can be produced by turbulence and (very effectively) by convection. Therefore in some statistical cloud schemes the production of variance is coupled to the activity of the turbulence and the convection scheme (in Harmonie this is the case when LOSIGMAS=TRUE and CMF_ CLOUD=STAT). However, this does not mean that in the absence of turbulent or convective activity the variance is zero. In nature other possible sources for variance are e.g. gravity waves, meso-scale circulations or large-scale advection. For situations without convective or turbulent activity, models therefore prescribe a small minimum value for the variance. However, in practice this minimum value can be too low sometimes leading to large areas with too little cloud fraction. To address this problem Geert Lenderink and Pier Siebesma came up with the idea of an extra variance term proportional to the saturation specific humidity. In principle this extra term adds characteristics of an old-fashioned relative humidity scheme to a statistical cloud scheme (personal communication Pier Siebesma). In 21

22 Wim de Rooy et.al. a relative humidity (RH) scheme cloud cover starts at a certain critical relative humidity value (RH c ) and then increases linearly with increasing RH. Formally the above mentioned correspondence between the extra variance term proportional to q sat and a RH-scheme can be written as follows: Q = q 1 q sat σ s NN = f (Q) q 1 q sat Assume σ s = αq sat = Q = = α 1 RH α 1 αq sat Where Q is the normalized distance to saturation curve, σ s is the standard deviation of s, NN is the cloud fraction, and α is a constant. Note that in the text we use the word variance instead of standard deviation (being the root of the variance). Now suppose that significant cloud cover starts at Q=-2, then RH c =1-2α. Here we assume that α=0.02 which corresponds to RH c =96%. Normally some turbulence will be present near the surface but at the same time q sat is relatively large in the lower part of the atmosphere, therefore the extra q sat variance contribution is multiplied by a function of height which reduces the influence in the lower troposphere (note that also RH-schemes are height-dependent). Details of this function are not discussed here. Results 13 May 28 We start with a case on the 13 th of May 28, a day with shallow convection in the area of The Netherlands and neighbouring countries, and large cloud fields over the North Sea. Although there are quite a few similarities between the cloud fields as shown by the satellite picture (Fig. 1) and the EDKF Harmonie run (Fig. 2a), two shortcomings stand out. Firstly, EDKF seems Fig. 1 Modis high resolution satellite picture for 13th of May 28, 131 UTC. The red square marks the location of Cabauw. 22

23 Wim de Rooy et.al. Fig. 2 Harmonie +36h forecasts valid for 13th of May 28, 12 UTC with a) EDKF (left panel) and b) EDMFm+ to have only a small part of the domain with small cloud fractions, less than 0.4, caused by the contribution of the DIRE option in the cloud scheme (i.e. a factor times the updraft fraction is added to the cloud fraction) and on the other hand areas which are totally overcast, so an all-or-nothing behavior. A better example of this typical behavior of EDKF will be presented later on. Secondly, a large part of the cloud field above the North Sea is disappeared. The latter cloud field is much better represented with EDMFm+ (Fig. 2b) so including the modified cloud scheme. Note that this cloud field above the North Sea also disappeared when EDMFm is run without the modification in the statistical cloud scheme. The modified cloud scheme does not result in excessive appearance of cloud fields (as happens when using LOSIGMAS=FALSE). The improvement in the on-off behavior in areas with shallow convection is probably the result of the different convection scheme as we will discuss in detail in the next case. The performance of the modified cloud scheme is also investigated for night times when convective activity over land is mostly absent and turbulent activity can be very small. Although less extreme, results for the night of the 13 th of May 28 shows similar improvements with the modified cloud scheme (not shown here). 9 April 2010 Apart from the impact of the chosen cloud scheme options, the next case is especially illustrative for the different behavior of the convection schemes in EDKF and EDMFm. Output is presented of 9 April We start with results from the KNMI testbed, i.e. Harmonie 1D results forced by the RACMO 3D model, valid for the Cabauw observation site. During the night this day starts with (stratus) cloud fields as indicated in Fig. 3 where the black dots represent measurements of the 23

24 Wim de Rooy et.al. Fig. 3 Contour plots of the cloud fraction from the 9 th of April 2010 at 0UTC until the 10 th of April 0UTC. The black dots represent lidar observations of the cloud base. Panel a) shows the results of RACMO, b) EDKF, and c) EDMFm+. Additionally panels b) and c) show the updraft lcl (solid line) and the updraft termination height (dashed line) lowest cloud base with a lidar instrument. It is unclear to what extent the night time clouds in the Racmo 3D run can be captured in the other models. In contrast to Harmonie, RACMO (Fig. 3a) uses a prognostic cloud scheme which can easily maintain clouds during the night while in a statistical cloud schemes the variance, and therewith the clouds, often disappear together with the dying of the turbulence and convection. Advection of cloud liquid water is not (yet) supported in the KNMI testbed but this might enable a closer match between the 1D models and the RACMO mother model for these kind of conditions. Nevertheless, the EDMFm version captures at least to a certain extent the (remains of the) observed night time clouds. More important is the following daytime period with shallow cumulus. The EDKF run (Fig. 3b) shows a several times observed problem with this version, namely overcast in combination with shallow convection whereas the observed cloud cover in these situations typically ranges from a few to approx. 30%. In contrast the EDMFm version (Fig. 3c) shows, after the disappearance of the remaining clouds from the night before, a more realistic shallow convection cloud profile. The relatively shallow, overcast, cloud layer in the EDKF run has a RH of 1% caused by the large amount of moisture transported into the cloud layer by the convection scheme as illustrated by Fig 4. Note that EDKF has only one updraft whereas for EDMFm+ only Fig. 4 Contour plots of the convective moisture flux (Wm -2 ) from the 9 th of April 2010 at 0UTC until the 10 th of April 0UTC. Panel a) shows the results of EDKF, and b) of EDMFm+ (only the moist updraft). Additionally the updraft lcl (solid line) and the updraft termination height (dashed line) are plotted. 24

25 Wim de Rooy et.al. Fig. Modis high resolution satellite picture for 9 th of April 2010, 1228 UTC the moist updraft is plotted. Therefore the moisture transport between EDKF and EDMFm can only be compared in the cloud layer (between the updraft lcl and the updraft termination height). Figure 4 shows the EDMFm+ scheme transporting moisture to the upper part of the cloud layer in the beginning of the shallow convection period. Due to the humid, favorable conditions for updrafts in the cloud layer in this period caused by the remains of the night time clouds (Fig. 3) the EDMFM convection scheme adjusts the mass flux profile to redistribute the moisture. Fig shows the satellite image for the 9 th of April 2010 case including the location of Cabauw. With a north westerly flow, shallow convection gradually develops above land, organized in streets of clouds. Figs 6a and 6b show the corresponding results of Harmonie EDKF and EDMFm+ respectively. These 3 D runs confirm the signal given by the testbed 1D output (Fig. 3). The EDKF version shows low cloud cover at the edge of a large cloud field, the result of the DIRE option, but most of The Netherlands is overcast. As mentioned above in the description of the tesbed results, the convection scheme fills up the cloud layer with moisture leading to full saturation and consequently 1% cloud cover. Just as in the EDKF run, the EDMFm+ run (Fig. 6b) starts the convection above land too soon and the cloud cover is again (although less than in EDKF) overestimated. Note that the 3D cloud cover output of Harmonie includes the effect of the cloud overlap function. To what extent the overestimation of the cloud cover in the EDMFm+ run is related to the overlap function has yet to be investigated. Further, Fig. 6 seems to indicate an improvement with the modified cloud scheme in cloud cover above the northern North Sea and above England. Noticeable is the very narrow line with low cloud fractions along the Dutch coast line in the EDMFm+ run (not visible on the satellite image Fig. ). Also this phenomenon should be further investigated. It is interesting to see how the situation develops three hours later. The (lower resolution) satellite picture (Fig. 7) resembles that of three hours before but the main difference is that the cloud structures became larger in the east and east of The Netherlands. It is encouraging to see that the Harmonie model (Figs. 8) is capable of reproducing large, resolved convective structures including the cloud free areas in between the convective systems, related to downdrafts. Again some improvement of the cloud field with the EDMFm+ option can be observed, e.g. north of The Netherlands, above sea. 2

26 Wim de Rooy et.al. Fig 6 Harmonie +24h forecasts valid for 9th of April 2010, 12 UTC with a) EDKF (left panel) and b) EDMFm+ (right panel) The on-off-behavior with the default EDKF-settings in comparison with EDMFm+ for Fig. 8 is supported by the histograms of the cloud fraction for this figure, shown in Fig. 9. Fig. 7 MSG satellite picture for 9 th of April 2010, 1 UTC 26

27 Wim de Rooy et.al. Fig 8 Harmonie +27h forecasts valid for 9th of April 2010, 1 UTC with a) EDKF (left panel) and b) EDMFm+ (right panel) Fig. 9 Histograms of the cloud fraction valid for 9 th of April 2010, 1 UTC (Fig. 8) with a) EDKF (left panel) and b) EDMFm+ (right panel) 27

28 Wim de Rooy et.al. Conclusions and discussion Results with the Harmonie model for two cases are presented containing 1D model output for the Cabauw super observation site and 3D output for an area around The Netherlands. These cases are representative for the general experiences at KNMI. Most striking is the all-or-nothing behavior with the default EDKF options. With the exception of areas with shallow convective activity, where sometimes small cloud cover is the result of the contribution of the DIRE option (a factor times the updraft fraction is added to the cloud cover) other cloud fractions are mostly 0 or 1. In the default EDKF configuration the variance for the cloud scheme only arises from the turbulent activity (for some reason multiplied by a factor 2) which results in too little variance and consequently the allor-nothing behavior. To address this problem, experiments are done with a full statistical cloud scheme, coupled to the turbulence and convection scheme, and the inclusion of a new extra variance term. This extra variance term adds characteristics of a RH-scheme to the statistical cloud scheme. Although the validation period with the latter new configuration is short and better tuning might be needed, the results are promising. The presented cases especially reveal improved representation of cloud fields above sea whereas the extra variance does not result in too large areas with non-zero cloud cover. Concerning the convection scheme, the results show that the EDKF scheme sometimes puts too much moisture in a too shallow cloud layer leading to unrealistic 1% cloud cover in shallow convective conditions (the cloud layer becomes fully saturated). The problem seems to be related to too large mass fluxes. Also the diurnal cycle of the cloud layer top is often not realistic. While in reality the top of the shallow cumulus layer normally increases gradually during the morning and early afternoon and decreases gradually in the late afternoon, the cloud layer top is often almost constant in EDKF runs. The EDMFm version uses for moist convection a very robust scheme based on ) and updated with ideas from 8). A RACMO version with the same convection scheme as EDMFm (called RACMO DualM TKE) already runs satisfactorily in the testbed for more than a year. Note that all Harmonie runs just became available in the testbed in March Unfortunately the EDMFm option as it is presently included in the official code does not use the correct logical settings nor it includes some bug fixes and updates as in the here presented EDMFm+ results. The updated EDMFm+ setting will probably become available in the next code upgrade. From then on experimenting with EDMFm+ in other countries then The Netherlands and exchanging the experiences with it between Hirlam Aladin communities is greatly encouraged. On a short term the daily testbed results will become accessible within a web browser for every one. Hopefully this can facilitate the research and development cooperation between Hirlam/Aladin members. 28

29 Wim de Rooy et.al. References 1) Neggers, Siebesma, Heus, 2010: Continuous single-column model evaluation at a permanent observational supersite. Submitted to BAMS 2) Pergaud, J., V Masson, S. Malardel, F. Couvreux, 29 : A parameterization of dry thermals and shallow cumili for mesoscale numerical weather prediction. Bound.-Layer 3) Siebesma, A.P., P. M. M. Soares, and J. Teixeira, 27: A combined eddy-diffusivity mass-flux approach for the convective boundary layer. J. Atmos. Sci., 64, ) Neggers, R.A.J., M. Köhler, A.C.M. Beljaars, 29: A dual mass flux framework for boundary layer convection. Part I: Transport. J. Atmos. Sci., 66, ) De Rooy, W. C. and A. P. Siebesma, 28: A simple parameterization for detrainment in shallow cumulus. Mon. Wea. Rev, 136, ) Chaboureau and Bechtold, 22: A simple cloud parameterization derived from cloud resolving model data: Diagnostic and prognostic applications. JAS Vol. 9, ) Soares, Miranda, Siebesma, Teixera, 24: An eddy-diffusivity/mass-flux parameterization for dry and shallow cumulus convection. QJRMS, ) De Rooy, W. C. and A. P. Siebesma, 2010: Analytical expressions for entrainment and detrainment in cumulus convection. Accepted for Quart. J. Roy. Met. Soc. 29

30 Per Undén Revised method of determination of the hybrid coordinate in HIRLAM by Per Undén SMHI, Norrköping, SWEDEN 1 Introduction The pressures of the hybrid model levels (in HIRLAM, ALADIN, IFS etc.) are defined as: p k+1/2 = A k+1/2 + B k+1/2 p s They need to be determined to cover the vertical depth of the model atmosphere as well as possible bearing in mind the number of levels that can be afforded. So first of all, some fairly equidistant resolution can be thought of. With a fairly limited number of levels, of or so, an equidistant resolution will give too poor resolution in the Atmospheric Boundary Layer (ABL) as well as in fronts and for the tropopause. At least in the ABL one should concentrate more levels closer to each other than higher up. Fronts are more difficult to cover since the vertical position varies widely. Also the tropopause varies in position, but it is in general between hpa and it can be desirable to concentrate layers also here. The hybrid coordinate system (η), above (Simmons and Burridge, 1981), is designed to go from a terrain-following σ system to pure pressure system in the stratosphere, where the flow is generally not depending on the surface. The value of B determines the degree of terrainfollowing coordinate at each level. So while we want to define the values of A and B to have the highest resolution in pressure (or height) in the areas of interest, like in the boundary layer, the variation of layer thickness needs to be smooth. First of all, since it is a coordinate it needs to be monotonous for all possible surface pressures. The other factor is that one does not want to introduce unnecessary vertical discretisation errors. In the earlier method used in HIRLAM, the values for the levels were determined through two polynomial representations; first for the values of η itself, then a second one for the values of B as a function of level number. For a description, see HIRLAM Newsletter No 41 (Undén and Gustafsson, 22). The η values were determined as to be as close as possible to some user defined target resolution, a specified list of level values that one would aim towards. The global (for all levels) polynomial fitting method encountered increased difficulties as demands for very high vertical resolution arose. It became more and more difficult to ensure that the resulting function (pressure as function of level number) was monotonous for all values of surface pressure in the model. Even after a lot of trial and error, the levels above mountains tended to become very thin (or even negative!) for 60 or more levels. In ALADIN (and ARPEGE) Pierre Bénard (24) developed a more tractable method which involves 30

31 Per Undén less trial and error and which is sounder and safer mathematically. It separates the problem into first a stretching function and then a hybridicity function. For different vertical regions functional expressions are defined with smooth transitions between those vertical regions. See the ALADIN (old) documentation pages under dynamics ( The method has been coded (there is also code available on the ALADIN pages above) and employed for a number of vertical resolutions for future HIRLAM and HARMONIE usage. The parameters for each resolution still need to be tuned and require some trial and error. The different vertical regions need to be defined (depth and number of layers in each). First, one needs to have some reasonable distribution between the different regions in the vertical. The fine tuning is then done in oder to get smooth second derivatives. Monotonocity is not guaranteed either, for any setting, but it is checked. The method from ALADIN is preferred for the future high vertical resolutions and in fact will be necessary since even after very hard trials the old HIRLAM method fails the monotonicity criterium for 1+ levels. 2 Definition of the vertical coordinate The following criteria are considered or aimed for in today s modelling in HIRLAM: Height of the lowest model level should be low enough for a) diagnostic purposes (near surface outputs for products and data assimilation), b) not to violate (too much) the assumption of constant flux in the Monin-Obukhov theory used for the surface layer and c) not be too low for instability and d) not too low to consider the flux aggregation from the ISBA tiled scheme. The latter is violated in most cases even at 30 m (1 m or more is thought to be appropriate for this but such high lowest model level is unrealistic). A lowering from the present HIRLAM level of about 28 m to 10 m seems appropriate from a) and b) above. ECMWF e.g. made this move several years ago (but on the other hand, their ABL scheme is not tiled). High resolution in the boundary layer. The description of the ABL and its processes is markedly improved (in 1D column models) when going to very high resolution (of a few 10:s of m). Troposphere with a good resolution and with smooth transitions from the ABL and to the stratosphere. A stratosphere with quite high resolution (in pressure) and with a high top for the purpose of satellite data assimilation and to some degree for the Upper Boundary condition (to be far away). A top level with a determined thickness, which determines the top (full) level pressure. So far HIRLAM has used 10 hpa for the upper top level, but it is inadequate for ATOVS and other satellite data with high peaking channels. The RTTOV code needs several levels above 10 hpa or it uses climatology for its background, which is clearly not satisfactory. It has been proposed to go to at least 1 hpa instead of the 10. (ECMWF has 0.01 hpa! (1 Pa) in the 91 level system). Xiaohua Yang et al. in the HIRLAM MG reviewed the issues of raising the top of model above 10 hpa. The radiation scheme would probably need some revision but above all, some control of high wind speeds which may occur at those high altitudes, is necessary (e.g. through increased horizontal diffusion). Also, an alternative to alleviate the RTTOV problem may be to supply additional ECMWF field data for the levels above the HIRLAM top. This has been planned in HIRLAM and should be possible but entails quite a bit of technical work. Thus, it has since 31

32 Per Undén become more relevant to derive new sets of model levels for the purpose of increased resolution as much as possible in the ABL (and in the lower troposphere and if possible also near the tropopause). The bottom of the atmosphere (a number of levels) uses pure terrain following levels whereas a number of the top levels uses pure pressure levels (the hybrid system, Simmons and Burridge (1981)). Apart from the top and bottom mentioned above, the total number of levels is of course decisive for the level of realism. First, a number of levels have to be assigned to be in the ABL and then a number of levels in the stratosphere. Remaining levels are in the troposphere. Then the pressure of the top of the ABL (or depth of the ABL) and the pressure of the bottom of the stratosphere have to be specified. The numbers of levels in each regime and the pressure limits are very important to tune since the overall distribution of levels has to be reasonable from e.g. smoothness criteria. The method devised by Pierre Bénard first defines a stretching function m(x) mapping model level numbers to m. x is jlev/nlev so x [0, 1]. The mapping is compressing in the ABL and in the stratosphere with a stretching in between. Certain forms of negative exponential functions are used in the ABL and in the stratosphere. The troposphere in between is fitted with a second order polynomial in order to have continuous values and derivatives in the transitions. Note that second order derivatives are not guaranteed to be small; therefore some experimentation and visualisation is required each time a new configuration is to be derived. The function m itself is nice and smooth, but plots of δm/δx should be used to tune the parameters (see also Fig 1). The exponential functions in the ABL and in the stratosphere have each an exponent which can and should be tuned for a final smoothness. (After having achieved a reasonable distribution of the levels in the three regimes as mentioned above). Then follows the hybridicity function h(y); where y = m(x) is shown in Fig. 2. The main purpose is to be == 0 at the top (y=0) and == 1 at surface (y=1) with a smooth transition in between. 1 'fort.11' 1 'fort.12' Figure 1: Stretching function m as function of relative level, x. Figure 2: Hybridicity function h as function of stretching m (of x) 32

33 Per Undén The application of m and h for pressure at a level (in terms of x = jlev/nlev ) is then: P(x) = P 0 [m(x) h(m(x))] + P s h(m(x)) where P 0 is the reference surface pressure ( hpa) and P s actual surface pressure. The values of A(x) and B(x) then readily follow from identifying the first term above with A and from the second term follows that B(x) = h(m(x)). 3 Proposed configurations and comparisons In connection with requests from the HIRLAM MG (earlier, in 27) for a new set of levels with in particular the top at 1 hpa, the new method was revived and tried for different numbers of levels. Somewhere in the range of levels seems reasonable for HIRLAM and compares quite well with ECMWF (apart from their much higher top). Hitherto the vertical resolution in HIRLAM has been somewhat lagging behind the one of the ECMWF operational model. The new method has been tried first for the historical resolutions of 31 and 40 levels and shown to give good vertical distribution of levels, even though the current aim in HIRLAM is to produce a new set of 70+ levels. Some HIRLAM members have experimented with increased vertical resolution, either from the old polynomial method or by using something from the ECMWF sets. A 71 level set was devised for the Working week in Copenhagen late August 27. (Earlier a 71 level set with the old method with a top at 3 hpa had also been produced but this one is not recommended due to very thin levels over orography and was only meant for technical purposes). As an illustration of the method Fig. 1 shows the stretching function m(x) for a 7 level version. One can see that the value of m increases very slowly at the top of the atmosphere (x=0) and at the bottom (x=1) with a more rapid increase in between and visibly very smooth transitions. The hybridicity function h(y); where y = m(x) is shown in Fig. 2. It goes from 0 at the top (y=0) to 1 at the surface (y=1) with a smooth transition in between. It is very important to check the distribution of levels in the vertical, most clearly by plotting the thickness of layers as function of x, i.e. the derivative of the pressure equation above. It is first shown (Figs. 3 and 4) for a surface pressure of hpa, i.e. it shows the effect of stretching (m(x)) only. This is almost always easy to get to look attractive. Also the old Reference 40 and 60 levels are plotted for comparison and the ECMWF 91 levels. It can be seen that the 7 level distribution is quite close to ECMWF 91 levels, with better (or in the ABL much better) resolution below hpa and slightly lower above hpa (by less than 10 %). Compared with the old Reference 40 levels it is a major resolution increase everywhere. The 60 levels that are used for some models actually have a higher resolution than the 7 levels in the troposphere but not in the ABL and not in the stratosphere. (See also below). The acid test is to plot the thicknesses over high orography where the hybridicity has a significant influence. It is never possible in a hybrid coordinate to get the same appearance as over sea level but one would like to avoid or minimise distorsions. Even though the derivative of P is continuous, the second derivative is not at all guaranteed and only after several adjustments of number of levels/ boundaries of the regimes and then the exponents, is it possible to get a relatively nice picture as in Fig. 4 for surface pressure of hpa. 33

34 Per Undén P (Pa) P (Pa) e dp (Pa) 1e dp (Pa) Figure 3: Level thicknesses for the earlier 40 and 60 levels and the new 7 levels and the ECMWF 91 levels. Surface pressure Pa and logarithmic scale. Figure 4: Fig.3 but linear scale. Figs. and 6 show again the HIRLAM 40 levels and the ECMWF 91 levels. It was for this picture, with hpa surface pressure, that the first attempt with 71 levels showed a coarser resolution than ECMWF all the way above the mid troposphere, e.g. at the jet level. Only by increasing the number of levels somewhat was it possible to obtain a quite similar picture as ECMWF. (The HIRLAM 60 levels used by some of us does NOT look nice in this picture and has some quirk near the bottom with an oscillation and now actually has a very similar resolution as the 40 level system. The old method spent all the effort increasing the resolution due to that oscillation around 40 hpa. In fact, for even lower surface pressure this derivative goes negative and the coordinate is invalid and the model will blow up.) Note that the 7 levels now are very close to the ECMWF levels in resolution. (Also please note that the levels above 1 hpa (to 0.01 hpa) are not shown). Table 1 compares a number of qualities between the configurations, old HIRLAM 31, 40 and 60 levels, the new 7 levels and the ECMWF 91 level system. It is clear that the 7 levels system has a high ABL resolution and is at least twice as good as P (Pa) P (Pa) e dp (Pa) dp (Pa) Figure : Level thicknesses for the earlier 40 and 60 levels and the new 7 levels and theecmwf 91 levels when the surface pressure is Pa (instead of 1013) and logarithmic Figure 6: Fig. but linear scale. 34

35 Per Undén 0 'fort.21' Figure 7: Cross section over a hpa hill for the 7 levels. the earlier configurations. It is a good increase below 10 m with 19 levels compared to 10 with 40 levels or just 13 for ECMWF. (AROME-v2 will have 27 levels below 30 m and 60 levels instead of 41). Finally, a cross section of full levels above an exponential mountain is plotted in Fig. 7 for that 7 level system. It shows a good regular behaviour for all pressures up to a minimum surface pressure of hpa (in ARPEGE the lowest pressure globally is 40 hpa and that is no problem either). For the meso-scale system (HARMONIE) it was proposed (in 27) to have a truncated version of the 7 level system presented here. Since the top levels are pure pressure levels, it would be possible to truncate at 10 hpa e.g. and interpolate a few levels down and then have identical hybrid coordinates as the 7 level system. This is what ECMWF appears to have done for the EPS 62 levels, but a plot of thickness does not look beautiful. There may be some reduction in interpolation errors of having identical levels, although surface pressures will anyway be different in the two models, host and guest models (so some vertical interpolation will take place anyway). It is instead better to derive a 60(+) level system which is quite close to the 7 levels except for the stratosphere. 4 Proposed new HIRLAM and HARMONIE set up with top at 10 hpa Following on from the thoughts in the last section, a setup with more than 60 levels would be desirable in HIRLAM and HARMONIE, to achieve a better ABL resolution at least. A 6 level set up has been derived which seems quite suitable from the distribution of levels ((Fig. 8 and 9) and in Table 1, below). As can be seen in the Table, the resolution is significantly better in the ABL than any of the others, and will have 33 levels below 30 m (more than the 27 in AROME v2). All the HIRLAM levels with top at 10 hpa will necessarily have a similarly poor resolution in the stratosphere, a mathematical necessity of finishing at 10 hpa only. It would be recommendable to return to the question of a few higher levels, too, even if only used as a spunge layer or even relaxed towards a coupling (host) model. 3

36 Per Undén 10 1 dp EC dp 60 dp 6 dp EC dp 60 dp 6 P (Pa) 0 1 P (Pa) 1e dp (Pa) 1e dp (Pa) Figure 8: Level thicknesses for the new 6 levels, the earlier 60 levels and the ECMWF 91 levels when the surface pressure is 1013 Pa and logarithmic scale. Figure 9: Level thicknesses for the new 6 levels, the earlier 60 levels and the ECMWF 91 levels when the surface pressure is 1013 Pa but linear scale and only up to 1 hpa. Table 1: Properties of the different model level configurations. Shows number of levels, height of lowest model level (m), number of levels below 1 m, thickness around 1 m, number of levels below 10 m, thickness in hpa at hpa, at 20 hpa and at 70 hpa and finally top pressure of full model level 1. conf Nlev Z nlev No <1m Dz 1m No <10m Dp Dp 20 Dp 70 Top HI HI HI HI EC HI This 6 level version has successfully been tested with HARMONIE at SMHI and also at DMI with HIRLAM and recent or ongoing tests show some improved scores for low level winds e.g. and no discernible negative effects. It is recommended that this version of levels is used for the future Reference, pending final evaluations. References Bénard, P., 24. Design of the hybrid vertical coordinate. ALADIN old documentation: Dynamics. Simmons, A.J. and Burridge, D.M., An energy and angular-momentum conserving vertical finitedifference scheme and hybrid vertical coordinates. Q.J.R. Met. Soc., 108, Undén, P. and Gustafsson, N. 22. Manipulations to determine the hybrid coordinate in HIRLAM. HIRLAM Newsletter No 41,

37 Huseyin Toros, Gertie Geetsema, Gerard Cats Evaluation of Hirlam and Harmonie precipitation forecasts for the Istanbul flash flood event of September 29 Huseyin Toros, Gertie Geertsema, Gerard Cats, Abstract The flash floods of 8 and 9 September 29 caused 31 deaths and substantial damage in the Istanbul metropolitan area. These events have been simulated with Hirlam and Harmonie. Both models give clear warning signs of the heavy precipitation in the area, but Hirlam in particular still underestimates the amounts. Due to its ability to advect hydrometeors over mountain ridges Harmonie predicts the large amounts with a better position than Hirlam. 1 Introduction In recent years several severe floods with high economical and social impact occurred in many parts of the world. In the beginning of September 29 a flash flood in and around the megacity Istanbul caused 31 deaths and resulted in material damage which was estimated to be of the order of 90 million dollars (see figure 1). The flash flood was caused by a period of two days of intense rainfall in the Thracian region. Under influence of a fast deepening low pressure system, warm moist air was transported from the Black Sea into the Thracian region, resulting in heavy showers and thunderstorms on the 8 th and 9 th of September. In this study we have performed Hirlam and Harmonie runs on domains centered at the Thracian region and Istanbul. The precipitation forecasts are evaluated against rainfall observations from some 28 meteorological stations in the region. Both models forecast heavy rainfall in the area, however the locations of the most intense rainfall are off by several tens of kilometers. 2 Observations 2.1 Synoptic Description On 6 September 29 a low pressure system with a central sea level pressure of 1 hpa is located Figure 1: Partially submerged cars are seen next to boats after heavy rains flooded Silivri on September 8 th, 29. Silivri is a town some 60 km from Istanbul Ataturk Airport. (Picture from Newspaper Todayszaman) 37

38 Huseyin Toros, Gertie Geetsema, Gerard Cats Figure 2: left panel: Weather chart showing the surface analysis for 8 September 29 UTC based on the operational Hirlam analysis (courtesy: KNMI operational service). The panel on the right shows the Meteosat 9 MSG IR image for the same time. over the Southwest of Anatolia and a high with a centre pressure of 102 hpa over central Europe. The low pressure system is deepening gradually whilst moving Northwards. On the 7 th of September two troughs develop at hpa. The surface analysis for 8 September is shown in figure Climate Istanbul is the business and cultural capital of Turkey and is home to about 13 million people (TUIK, 2010). It has historically been vulnerable to natural disasters. The climate of Istanbul can be characterized as a transition between Mediterranean and Temperate. In the summer months the climate is generally warm and humid with very little rain, whereas the winter months can be cold and wet with some snow. The spring and autumn seasons are mild. Istanbul covers a large area and has a complex topography. The average annual precipitation in the Northwestern part of Turkey, the Marmara Region, is approximately 7 mm with a range from 40 mm to 870 mm. 2.3 Observations Hourly precipitation data at the 28 stations shown in figure 3 were obtained from the Turkish State Meteorological Service (TSMS) and the Istanbul Metropolitan Municipality (AKOM). Table 2 shows that already on 7 September large amounts of rain had fallen in the area. However this rain fell outside the catchment areas related to the flash flood events. To put the amounts of observed precipitation into perspective we mention that the mean annual precipitation in this area is slightly less than 7 mm. For the 28 stations the average observed precipitation on the two days 8 and 9 September is 84 mm, more than 10% of the mean annual total precipitation. The two day accumulated observed precipitation at station Catalca (id 11) is even 240 mm, approximately one third of the average yearly precipitation. Figure 1 shows the effect of such heavy rainfall. The picture is taken at Selimpasa, Silivri coast, which is indicated in figure 3. The amount of rainfall was so high that it exceeded the capacity of Ikitelli Ayamama Creek, an important drainage canal for the Istanbul metropolitan area. This creek is also indicated in figure 3. On 8 September heavy rain falling uninterrupted since the early hours affected the region 0 km west of Istanbul causing flood disasters in many towns. Shop windows crashed, flood water entered businesses and homes, many cars drifted into the sea. Figure 3 indicates the catchment area and the 38

39 Huseyin Toros, Gertie Geetsema, Gerard Cats Figure 3: The top panel shows a map with the precipitation statons in the Northwestern part of Turkey (the Marmara Region). The station numbers correspond to those in table 2, the numbering increases from West to East. The rectangles in the bottom panels show the areas of flooding for 8 (left) and 9 (right) September. The polygons in the top panel indicate the relevant catchment areas. area of flooding. The lower left panel shows Selimpasa, Silivri coast. Silivri is a district of Istanbul Province along the Sea of Marmara, outside of metropolitan Istanbul, containing many holiday and weekend homes for residents of the city. On September 9 the Ayamama Creek overflowed which resulted in roads almost completely covered by water. The flooded region, Ikitelli, in the Istanbul urban area, is a district with much traffic and industry. Figure 3 also shows the catchment and flooding areas for this event. 3 Model setup The flood events on September 8-9 were examined using Hirlam and Harmonie. The models were run at ECMWF using the default settings. The number of gridpoints is given in table 1. The model domains are shown in figure 4. The Hirlam run is initialized at UTC on 6 September with 48-hours forecasts every 12 hours. The Harmonie run is started at UTC on 6 September with analyses every six hours. The forecast length is 24 hours. Both models are nested within ECMWF analyses. Table 1: Details of the model setup Model feature Hirlam Harmonie PE Hydrostatic Non-hydrostatic Resolution 11 km 2. km Gridpoints 438x4 389x389 (189x189) 4 Model simulations The precipitation forecasts for 7 September are also shown in figure 4. According to the Hirlam forecast some 10 to 20 mm can be expected in the region Northwest of Istanbul. Harmonie shows peak values of 70 to 1 mm, which compares well to the observed values. For example the average daysum of stations 4 and is 77 mm. Visual comparison of the model forecasts with the radar information given in figure confirms these conclusions. The flash floods are mainly due to heavy rainfall on September 8-9. The weather situation leading to these large rainfalls is described by Schipper and Erturk (29). They found that the upper level low over central Turkey results in the advection of moist unstable air from the Black Sea. Subsequent 39

40 HIRLAM Newsletter no 6, November 2010 Huseyin Toros, Gertie Geetsema, Gerard Cats Figure 4: Accumulated 0-24 hours precipitation forecast from 7 September UTC until 8 September UTC. The plotted area shows the model domain. Left panel: Hirlam, right panel: Harmonie. topographic uplift then results in numerous rounds of showers and thunderstorms in the Marmara Region (west of the Black Sea). The low temperatures found in the upper layers of the cold core low and its contrast to the warm surface causes CB s to grow enthousiastically, producing lots of rain. To achieve accurate short-term precipitaton forecasts in complex terrain, such as in the greater Istanbul Figure : SLI_R radar images are given every six hours starting at 7 September UTC. The Doppler weather radar in Istanbul is operated by TSMS. 40

41 Huseyin Toros, Gertie Geetsema, Gerard Cats area, mesoscale models must include an accurate treatment of the precipitation process and full dynamical interactions driven by the fine-scale variability of topography and atmospheric conditions. During the heavy rain event (7-10 September), 270 mm has been measured in Catalca in the west of Istanbul. The precipitation on 8 and 9 September averaged over the 28 stations is comparable to long term monthly averages for that region. 4.1 September 8 Figure 8 shows observations of rainfall on 8 September. The rain-gauge observations show large spatial variability between Northwest and Southwest of Marmara region. The highest amounts of rainfall are measured in Catalca (204 mm), Bandirma (122 mm) and Gonen (109 mm). However Bandirma and Gonen are located South of the Marmara Sea and not in the catchment area of the area which was flooded on 8 September. The precipitation value of 49 mm measured at station Terkos (id 14) is most probably a severe underestimation of the true value, because the station did not supply data from 08: until 23: UTC. The radar image in figure 9 suggests that the rainfall at Terkos may well have been as heavy as at station Catalca (id 11), 204 mm. The average over all stations is 43 mm. The average over the Silivri flooding area stations (the blue coded stations in table 2) is 81 mm. The bulk of the heavy rain is offshore and over Bandirma (id 9) along with a significant spread along the Southwest of Marmura Sea. The Hirlam forecast shows two small areas with peak values in the range of 70-1 mm, see figure 8. The peak value area in the middle of the picture is some 20 km Northwest of Catalca (id 11, see fig. 6), at the windside of the model orography (fig. 7). Figure 6: Hirlam grid point values showing the 24-hours precipitation forecast for 8 September (valid at 9 September UT). Figure 7: Hirlam (top) and Harmonie (down) orography. 41

42 Huseyin Toros, Gertie Geetsema, Gerard Cats Table 2: The table gives the station id, name, longitude, latitude, altitude (m), daily total observed precipitation (obs) and the models 24-hour accumulated precipitation forecasts for Hirlam (hir) and Harmonie (har), all in mm. The colors indicate stations which are in the catchment area relevant for 8 (blue) and 9 September (green station names). The * in column 7 indicates that station Terkos (id 14) did not register any data from 08: to 23:, therefore the number given is a 9-hours precipitation sum. The values for Hirlam and Harmonie have been interpolated between grid point values. 7 Sep 8 September 9 September id station name lon lat alt obs obs hir har obs hir har 1 Ipsala Uzunkopru Malkara Kirklarel Luleburga Tekirdag Gonen Corlu Bandirma Canta Catalca Kamiloba Hadimkoy Terkos Olimpiyat Florya Aksaray Akom Kumkoy Sariyer Buyukada Cavusbasi Kartal Samandira Omerli Sile Kocaeli Sakarya Average The Southern peak value area collocates with station Bandirma. Along the coastline Northwest of Istanbul Hirlam has clear warning sings for large precipitation values, however underestimates the observed values, and has the precipation too much to the North of the catchment area. The Harmonie forecast is plotted in the bottom panel of figure 8. Harmonie forecasts daysum up to almost 4 mm, which is quite excessive in comparison with the annual precipitation of 7±10 mm, but is in fact only a factor 2 larger than the observed 204 mm in station Catalca. Concentrating on the Istanbul area we see that the peak values of Harmonie are located to the South of the observed peak values. It is a common feature that Hirlam forecasts the precipitation on the windside of the hill range. Its explanation lies in the fact that in Hirlam precipitation falls vertically. Harmonie, on the other hand, 42

43 Huseyin Toros, Gertie Geetsema, Gerard Cats Figure 8: The 24-hour rainfall totals (mm) in Istanbul and its surroundings on 8 September 29. The right panel shows the measured precipitation, the top left panel the 24-hours forecast from Hirlam and the bottom panel the 24-forecast from Harmonie. The * in the top panel indicates that site Terkos did not register any data from 08: to 23:. Figure 9: SLI_R radar images are given every six hours starting at 8 September UTC. The Doppler weather radar in Istanbul is operated by TSMS. 43

44 HIRLAM Newsletter no 6, November 2010 Huseyin Toros, Gertie Geetsema, Gerard Cats Figure 10: Same as figure 8 for 9 September 29. Figure 11: Same as figure 9 for 9 September