To improve severe weather forecasting, KNMI has started a special programme in 2005. A number of research projects are included, like KOUW (Probabilistic forecasts of thunderstorms for the purpose of issuing a weather alarm) and MESOMOD1).
Other, non-research projects of the programme are aimed at better tuning of severe weather information to the needs and wishes of the user, at a clear and consistent communication to the user and media, and at a more dedicated role of the forecaster.
In this highlight the most important results of the KOUW project are presented. Because the skill of severe thunderstorm warnings was unsatisfactory, the KOUW project was initiated. The goal of this project was to develop a probabilistic forecast system for (severe) thunderstorms to be used by forecasters as a tool to decide whether a weather alarm should be issued.
In this project the technique of Model Output Statistics (MOS)2) has been used to derive logistic regression2) equations for the (conditional) probability of (severe) thunderstorms in the warm half-year (from mid-April to mid-October) in the Netherlands. The MOS technique consists of determining a statistical relationship between a predictand (i.e. the occurrence of a thunderstorm in this
case) and predictors from numerical weather prediction (NWP) model forecasts and possibly from observations. For 12 regions of about 90x80 km2 each (Figure 1) and for forecast projections up to 12 hours in advance (with 6-hour periods), we have developed these equations using several potential predictor sources. These sources not only consisted of combined (postprocessed) output from two NWP models, as in our previous study3), but also contained an ensemble of 18 members of advected radar and lightning data for the 0-6 h projections4). As a source for the predictands, reprocessed lightning data from the SAFIR network5,6) have been used. The system was made pre-operational at KNMI in the spring of 2006 and produces an update every 3 hours during the warm half-year.
Two predictands are defined, based on (reprocessed) SAFIR5,6) lightning data. Herewith, both horizontal and vertical lightning discharges are taken into account. One predictand is defined as the probability of a thunderstorm (i.e. > 1 discharge in a 6-h period in a 90x80 km2 region). The other predictand is defined as the conditional probability of a severe thunderstorm (with maximum intensity thresholds of 50, 100 and 200 discharges/ 5 min. in a 6-h period in a 90x80 km2 region) under the condition that > 1 discharge will be detected in the same 6-h period in the same region. The two highest thresholds are only available for the 12-18, 15-21 and 18-00 UTC periods. The absolute probability of a severe thunderstorm can be computed by multiplying the conditional probability of a severe thunderstorm by the probability of a thunderstorm (> 1 discharge). Note that the predictand definition for severe thunderstorms is different from the one used in the operational severe thunderstorm forecast system3).
The current weather alarm criterion for severe thunderstorms (≥ 500 discharges/ 5 min. in any area of 50×50 km2) is met only twice a year on average. For such a rare event it is impossible to derive skilful statistical equations. Therefore, we have investigated what the maximum intensity threshold is that renders a skilful system. This has led to the above mentioned predictand definitions. Not only the intensity thresholds in the KOUW-system are lower than the weather alarm threshold, but the area size is also a factor of 3 larger. The area size is larger, because the thunderstorm frequency would otherwise become so low (in the night and morning) that there would be too few thunderstorm cases left to derive statistical equations. On the other hand, the 90x80 km2 regions in the KOUW-system are specified a priori, while the weather alarm criterion is defined for any area of 50×50 km2.
New compared to the operational thunderstorm forecast system3) is the use of an ensemble of 18 members of advected radar and lightning data as potential predictor sources for the 0-6 h projections4). The lightning and radar images of 02.40, 05.40, 08.40, 11.40, 14.40, 17.40, 20.40 or 23.40 UTC have been used as initial conditions. Subsequently, both the lightning and the radar image are advected using the HIRLAM 700 hPa wind vectors together with vectors computed from previous radar images. Apart from these basic vectors, also vectors that are 25% longer and 25% shorter have been used and vectors whose direction deviates by +10 and -10 degrees, respectively. This leads to a total number of 18 ensemble members.
The remaining potential predictor sets consist of 17 thunderstorm indices, computed from the 22-km HIRLAM reforecasting dataset, ECMWF (postprocessed) direct model output (resolution: 0.5 degree) and the (co)sine of the day of the year. The dataset used to develop the application consists of July 2002 to July 2005 data (warm half-years only) with independent verification datasets consisting of July to mid-October 2005 and mid-April to mid-October 2006 data, respectively. Because the severe thunderstorm sample size is too small for each region separately, all 12 regions have been pooled in the severe thunderstorm forecast system.
After a selection process4), the most important predictors in the thunderstorm forecast system turned out to be the percentage of the total number of advection ensemble members with , ≥ 4 discharges (0-6 h projections only), the ECMWF 6-h convective precipitation sum, the HIRLAM Jefferson index, the HIRLAM CAPE of the most unstable level and the HIRLAM Boyden index (see Table 1 for definitions of the thunderstorm indices). The most important predictors in the severe thunderstorm forecast system turned out to be the HIRLAM Bradbury index, a number of predictors from the ensemble of advected lightning data (0-6 h projections only), the ECMWF 3-h convective precipitation sum and the HIRLAM Jefferson index. In the logistic regression2) equations for the three different thresholds (50, 100 and 200 discharges/ 5 min.) the same predictors are used, but of course with different regression coefficients. The equations contain minimally 2 and maximally 5 predictors. The maximum number of predictors has been set to 5, because more than 5 predictors often appeared to result in overfitting.
Figure 1a shows an example of a probabilistic forecast of severe thunderstorms for 15-21 UTC on June 25, 2006 and Figure 1b shows the observed maximum 5-min. lightning intensity for the same period. In the morning the system already showed high probabilities of severe thunderstorms, especially in the south-eastern regions (Fig. 1a). It is found that the 200 discharges/ 5 min. threshold was exceeded in 3 regions (Fig. 1b). Inside the Netherlands the weather alarm criterion has not been reached, but just outside (within the most south-eastern region in Fig. 1b) it has. In fact this was the only case in 2006 in which the weather alarm criterion was met anywhere within the total area of the 12 regions. Of course, (probability) forecasts cannot be verified using only one single case, so objective verification results for the independent datasets are presented here.
We conclude from the verification results for 2006 (Figure 2) that the overall skill of the MOS thunderstorm forecast system (> 1 discharge) is good compared to the 2000-2004 climatology. The average Brier skill score (BSS)2) of the 6 land regions (EMN, MMS, EMS, WXS, MXS and EXS)3) shows a clear diurnal cycle with the highest skill in the afternoon (12 and 15 UTC) and evening (18 UTC). The average BSS2) of the 6 coastal regions (WXN, MXN, EXN, WMN, MMN and WMS)3) shows a diurnal cycle as well, but with a smaller amplitude and a phase shift of approximately 12 h. This can be partly explained by the smaller and different diurnal cycle in the occurrence of thunderstorms in the coastal regions compared to the land regions. The Brier skill scores for the 6-12 h forecasts (Figure 2b) are generally smaller than those for the 0-6 h forecasts (Figure 2a), as expected. Apart from the fact that the skill of a forecast system decreases with increasing forecast projections, the loss of the most important predictor for the 0-6 h projections (i.e. the percentage of the total number of advection ensemble members with ≥ 4 discharges) is expected to play a role as well. Finally, the Brier skill scores of this new thunderstorm forecast system are generally higher than those of the operational system3) (not shown).
Severe thunderstorms were relatively rare in the warm half-year of 2006. Because the period July to mid-October 2005, in which their frequency was about the same as in the 2000-2004 period, was more representative, we have verified the MOS system for severe thunderstorms over both periods. In Figure 3 reliability diagrams2) are shown for the 0-6 h forecasts of 5 runs of the MOS severe thunderstorm forecast system based on the lowest threshold of 50 discharges/ 5 min. Figure 4 shows reliability diagrams based on the higher threshold 0f 100 discharges/ 5 min. The severe thunderstorm forecast system shows the highest skill in the evening (15-21 UTC) with the highest skill for the lowest threshold, as expected. The reliability diagrams for the period July to mid-October 2005 show a better overall skill than the diagrams for the warm half-year of 20067). Factors that may have contributed to the lesser skill in the warm half-year of 2006 are sampling effects, the increased ECMWF model resolution starting February 2006 and/or the low frequencies of severe thunderstorms. Nevertheless, the evaluation of the system by forecasters during the period May 24 2006 to mid-October 2006 was positive7). Apparently, a forecast system can be much more useful in practice than the (harsh) objective verification results might suggest. We conclude from the verification results shown here (Figures 3 and 4) that the overall skill of the severe thunderstorm forecast system is reasonable compared to the 2000-2004 climatology, at least for the thresholds of 50 and 100 discharges/ 5 min. Conclusions for the higher threshold of 200 discharges/ 5 min. can only be reached when a dataset with more severe thunderstorm cases will become available. Finally, we would like to stress that the verification results for the independent dataset during the development were good for all thresholds (not shown).
The pre-operational KOUW system will be extended to include the 12-48 h projections. Subsequently, this system will replace the operational (severe) thunderstorm forecast system3). Future developments in this MOS system may consist of even more extreme criteria for severe thunderstorms4) and the inclusion of advected MSG data as a potential predictor source for the 0-6 h projections.