How much can we expect from forecasting systems in the future? European floods of 2013 – revisited

How much can we expect from forecasting systems in the future? European floods of 2013 – revisited

Contributed by Fredrik Wetterhall

In early June 2013, the Northern part of the Alps where hit with extreme rainfall that led to extensive flooding in Central Europe, with a death toll of 25 people and overall economic losses of €12bn. A classical synoptic weather pattern with a deep quasi-stationary low pressure system brought moist air towards the Northern side of the Alps. Orographic lifting led to very intense  precipitation over a large area where the soils were already saturated due to earlier wet weather in May.

Figure 1.
Figure 1. The total precipitation over 72 hours (31 May 06UTC-3 June 06UTC 2013) for ECMWFs operational high resolution model (left) and the observed precipitation generated from EFAS gridded observations (right)

The location of the precipitation was accurately predicted by most global numerical weather prediction models (NWPs),  however the magnitude  was widely underestimated (Fig.1). The discharge in the affected areas was well predicted by EFAS, and a total of 14 flood warnings were sent out prior to the event (see example for Passau in Fig 2). The magnitude of the flooding was nevertheless underestimated, to a large extent as a consequence of the underestimation in precipitation.

Figure 2. EFAS forecast for Passau on the 30 May 2013
Figure 2. EFAS forecast for Passau on the 30 May 2013. The lines and boxplots denote the different driving forecasts used.

The current ECMWF high resolution forecast runs on a grid size of about 16 km, and on that scale the synoptic circulation is well resolved, but intense precipitation associated with deep convection is parameterised. An increase in resolution as well as improving the model physics would potentially lead to improvements in the modelling of precipitation for extreme event. To assess the effect of resolution as well as  model physics on precipitation, we ran a number of model experiments over the event to see which improvement would be most beneficial.

Experiments with increased resolution and improved cloud physics

The first hypothesis was that increased resolution would also mean a better described orography, whin in turn would increase the orographically enhanced precipitation. We therefore reran our latest model version (40R1) with resolutions T319, T639, T1279 and T2047, which translates to 64, 32, 16 and 10 km respectively. The high resolution model currently runs at T1279, but will increase to T2047 in 2015. We also ran an experimental model version at resolutoin T3999, which is about 5km. That is the resolution which is in the so called “grey zone”where the relationship between parameterized deep convection and explicit deep convection on the grid scale becomes less clear.

Not only resolution is important for precipitation, but also cloud microphysics, especially the formation of precipitation. Therefore the current formulation proposed by Sundquist (SQ, Sundquuist et al, 1978) was compared with a new scheme proposed by Khairoutdinov and Kogan (KK, 2000) for the resolutions T1279 (16 km) and T3999 (5 km). An experiment allowing deep conversion was also tested for the highest resolution. For a full description of the experimental setup, see Haiden et al. (2014).

Figure. Cumulative precipitation
Figure3. Accumulated precipitation over the alpine box (47-48N, 10-14 W) over 72h starting from 31 May 00UTC. The left figure shows the experiments with different resolutions, and the right figure with different cloud physics schemes for the T1279 and T3999 resolutions. “noconv” indicates when the deep convection scheme was turned off. The black line denotes the precipitation calculated from a gridded observational dataset.

The experiments were validated over a small box over the alps (47N-48N, 10W-14W) where most of the rain fell. This location was also chosen to fully investigate the effect of increased resolution on the orographically enhanced precipitation. Figure 3, left panel shows the increase in precipitation with increase in resolution, which can be attributed to a more detailed orography. However, there is still a substantial gap to the observed precipitation of the event. The impact of the new cloud physics is shown on the right hand side of Figure 3. The current operational model with the inclusion of the KK scheme (light green curve) shows a substantial improvement. The most impact of the improved cloud physics is with the highest model resolution. The KK scheme is better than the SQ scheme, but a tuning of the latter scheme gives an even higher precipitation intensity. Finally, the experiment allowing deep convection brings the precipitation very close to the observed intensities. The T3999 with deep convection also improved the spatial distribution of precipitation (not shown).

Hydrological impacts

The discharge over the Alpine region were also better modelled, and Figure 4 shows how the improvements in precipitation translated into modelled discharge for Passau, which was one of the most severely hit locations. Even though the flood peak becomes closer to the hydrograph modelled with observed precipitation, it is still far off. However, this is result for one station for a single model run, and to really assess the impact of the model improvements, more cases are needed. It also highlights the importance of ensemble forecasting for specific watersheds; even it the precipitation over a large region is well captured, spatial variations in the precipitation field can mean the difference between a flood or not.

Forecasted discharge levels with two different model resolutions (T1279 and T3999) for the station Passau at the river Danube. The black line denotes the simulated discharge using observed precipitation.
Figure 4. Forecasted discharge levels with two different model resolutions (T1279 and T3999) for the station Passau at the river Danube, with and without the improved cloud physics (deep convectoin on for the T3999). The black line denotes the simulated discharge using observed precipitation.

These experiments have shown that improvements in resolution and model physics will lead to also better forecasts of extreme events and we can expect that this will lead to better forecasts in the future. However, this is one case study, and one caveat of tuning the model for a specific event is that it can yield unwanted effects in other situations. More studies on similar events are called for to show the numerical weather modeling community the need to improve forecasts of extreme events.


Haiden, T., Magnusson, l., Tsonevsky, I., Wetterhall, F., Alfieri, L., Pappenberger, F., de Rosnay, P., Muñoz-Sabater, J., Balsamo, G., Albergel, C., Forbes, R., Hewson, T., Malardel, S., Richardson, D., 2014, ECMWF forecast performance during the June 2013 flood in Central Europe, ECMWF Technical Memoranda, 723, 34 pp, Reading, United Kingdom.

Khairoutdinov, M., and Y. Kogan, 2000: A new cloud physics parameterization in a large-eddy simulation model of marine stratocumulus. Mon. Wea. Rev., 128, 229-243.

Sundqvist, H., 1978: A parameterization scheme for non-convective condensation including prediction of cloud water content. Quart. J. Roy. Meteor. Soc., 104, 677-690.

6 thoughts on “How much can we expect from forecasting systems in the future? European floods of 2013 – revisited

  1. A very intersting result and very similar to what we found with the MMEFS ( for a number of years with NOAA/NWS. But I can not help wondering about the hydrologic models themselves, because – if I understand the inset of Figure 2 correctly – it appears that the timing of the hydrologic ensembles is about 48-hours too slow; unless, this is the result of the timing of the NWP ensemble precipitation? Could it be that the hydrologic models need improved calibration? Certainly, if the deterministic QPF is underpredicted, the hydrologic models can not be expected to perform adequately. But one would think that the ensemble precipitation might capture the timing of the event better than by being 48-hours too slow. Just a thought…

  2. Thomas, the lines in fig 2 are all forecasts, but with different NWPs. The black line is the DWD forecast from the German weather service, and the box plot is the ECWMF ensemble forecast. In this particular case and catchment, the DWD model had the timing better than ECMWF, but this is one case so on conclusions in terms of general timing can be drawn.

    The question regarding calibration of the hydrological models is more complex. You can choose the best model for your catchment and calibrate it ti be a very good prediction instrument. In that case it would be the NWPs that would infer the timing problem. In a pan-European system like EFAS it is not so easy, since the model has to perform on a European scale. Therefore the calibration cannot be optimized for each singe location.

    In this study the timing of the NWP was generally improved with resolution and cloud physics, as you can see in figure 3. It is also seen to some extent in Figure 4, however the results can differ from cachment to catchment.

  3. Hi Frederik, very nice post!
    I was wondering if you tested or maybe is planning to do some kind of “backward” experiments, following this line: a) What was the NWP necessary to generate this observed flow peak in the hydrological model? b) How would it possible to happen with the observed conditions you have at that time? c) Is this answer feasible? Or there are things that are not being considered in the system that must be improved for it to be feasible? (for example, maybe there are hydrological issues that are not being considered, or maybe we need to consider other uncertainties).

  4. Hi Fernando,
    Very interesting idea. There will be afollow-up paper on the hydrology, and in that we will discuss more on the soil water and the hydrological response to the forecast. In the post I am focussing on precipitation over the event, but the antecedent conditions were very important in this case and you could calculate the water “needed” to recreate the event. However, I think that you would need to go to the 1-2 km scale in your NWP to really be able to model the convection and orographic. I am here focussing on the coarser global models, but it should also be said that limited area models could pick up the details of this event much better (see Yan, X., C. Wittmann, and F. Meier, 2014: AROME in Austria. ALADIN-HIRLAM Newsletter No. 2, 27-32)

  5. Hi Fredrik,
    Interesting analysis — it’s promising to see that local precipitation intensities are enhanced by the finer resolution runs and new parameterizations. I expect that even with such measures, precipitation in the largest extreme events will always tend to be somewhat underforecast, and I wonder how the current operational run with post-processing (downscaling/calibration) would have performed, relative to the other enhancements? We know from the work of Tom Hamill and others that it can boost performance, particularly for problems like bias (if systematic). Do the current runs have similar potential predictability, despite their low bias? Perhaps in the future paper, one baseline can be the current operational setup, and another can be the setup with post-processing. -Andy

    1. Andy,
      The EFAS system does provide post-processing of the modelled hydrograph using bayesian modelling averaging techniques for the points where observed discharge is available in real time.

      Another possibility is calibration of the forecast using hindcasts. In a recent paper, Alfieri et al compared the EFAS forecasts over the event with a runoff index which was calibrated against its hindcast. The runoff index does not contain any routing component more than the concentration time of a sub basin which means that it cannot be used on larger basins where routing becomes important. However, it is very useful for small basins as an indicator of flash floods. The index did perform well for the event and indicated higher probabilities of an extreme event than the EFAS forecasts.

      Using post-processing or bias correction of the results from the current study will not be possible since we were only able t run a few experiments and do not have enough simulations to be able to apply post-processing methods.

      Alfieri, L., Pappenberger, F. and Wetterhall, F.: The extreme runoff index for flood early warning in Europe, Nat Hazards Earth Syst Sci, 14(6), 1505–1515, doi:10.5194/nhess-14-1505-2014, 2014.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.