SEARCH

SEARCH BY CITATION

Keywords:

  • forecast;
  • probabilistic forecast;
  • ensemble;
  • Venice flooding;
  • historical floods

Abstract

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

The accuracy of deterministic and probabilistic forecasts of storms in the Adriatic Sea that lead to the flooding of Venice is discussed. We consider ECMWF state-of-the-art high-resolution single deterministic and lower-resolution ensemble-based forecasts of meteorological and sea states (waves) for five storms that affected Venice between 1966 and 2008. Notwithstanding the complicated local orographic situation, it is shown that ECMWF single, deterministic forecasts provide accurate information up to 3–4 days in advance. This range is further extended to between 4 and 6 days if ensemble-based, probabilistic forecasts are considered. The assessment of the quality of Ensemble Prediction System (EPS) probabilistic forecasts during the winters of 2008–2009 and 2009–2010 over the Adriatic and the Mediterranean Seas, and the North Atlantic Ocean, is also discussed to provide a proper statistical evaluation of the accuracy of EPS-based probabilistic forecasts of the wind over the sea. Average results indicate that EPS probabilistic forecasts over these areas are skilful for the whole forecast range considered in this study. Copyright © 2011 Royal Meteorological Society


1. Introduction

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

The quality of meteorological forecasts has been steadily improving with time. This is due to the larger availability of data, model improvements (description of the physical processes, numerical methods and resolution) and advances in data assimilation made possible thanks to ever-increasing computer power. As an example, today the deterministic forecast of the 500 hPa geopotential height issued by the European Centre for Medium-Range Weather Forecasts (Reading, UK, henceforth ECMWF) crosses the 60% anomaly correlation coefficient at about forecast day 7.75, while in 1980 this happened at around day 5.2. This suggests an average improvement of about 1 day per decade in the prediction of the synoptic scales represented by the 500 hPa geopotential height (Palmer et al., 2007).

Despite the continuous increase in skill, single forecasts such as the ones generated using the ECMWF single deterministic (henceforth DET) system, which now runs with a spectral triangular truncation T1279 (equivalent to a grid spacing of ∼16 km at midlatitudes) and 91 vertical levels, sometimes fail to be accurate. Although this happens rather rarely, it is often associated with fast developing, small-scale features that can cause severe disruption and damage at local levels. To address this weakness, and in general to enrich weather forecasts by providing a more complete description of the range of possible future weather scenarios and an estimate of the related reliability (Buizza, 2008), the past two decades have seen the development and operational implementation of ensemble-based probabilistic systems. At ECMWF, the Ensemble Prediction System (EPS) has been part of the operational suite of models since 1992 (Buizza and Palmer, 1995; Molteni et al., 1996). Since then, it has been providing forecasters with an estimate of the full probability distribution of forecast states. For example, a probabilistic forecast can be used to assess the probability of occurrence of extreme sea states that could lead to severe losses and damage. Parallel to the deterministic approach, the past decade has also seen the skill of the EPS continuously improving. Taking as reference the skill of the probabilistic prediction of the 500 hPa geopotential height or the 850 hPa temperature, the gain has been about 1.5 days per decade (Palmer et al., 2007)

The implementation in 1992 of the ensemble-based prediction system at ECMWF and at the National Centers for Environmental Prediction (NCEP, Washington) was rapidly followed by a similar action at many other meteo-oceanographic forecast centres, such as Australia, Brazil, Canada, UK Met Office, Japan, Korea, and NCEP in the USA. Park et al. (2008) give a review of all the global ensemble systems operational in 2007. Recent work has also extended the use of EPS to a variety of applications, such as the risk management of the probability of flooding (e.g. Gouweleeuw et al., 2005; Schaake et al., 2007) or energy demand (Taylor and Buizza, 2003). In particular, practical experience (e.g. Buizza and Chessa, 2002; Buizza and Hollingsworth, 2002) has shown that, for extreme events, the EPS approach can provide valuable information between 12 and 24 hours earlier than single DET forecasts. The range of useful forecasts has been further extended with the coupling between the meteorological and wave (WAM) models both in the deterministic and ensemble approaches (Sætra and Bidlot, 2004).

These results are typical of open spaces (i.e. of areas without any strong boundary constraints such as valleys or enclosed basins), especially when we focus on the sea surface. While indeed the quality of medium-range wave forecasts is very high in the oceans (e.g. Janssen, 2008), this is frequently not the case in enclosed seas or inner coastal areas. In these latter cases, apart from other problems such as those linked to orography, even a small shift of the storm may change substantially the local wave conditions because of the critical dependence on fetch. On the other hand, this is exactly the situation when it is worthwhile to explore how much the output will change by suitably perturbing the analysis, i.e. the conditions from which the ensemble forecasts start. From the distribution of the resulting possibilities one can derive how critical, in other words how reliable, the forecast can be and have an estimate of the probability of events to worry about. It is therefore of interest to explore the predictability of both the deterministic and ensemble prediction systems in the inner seas.

This work discusses the use of single deterministic and of probabilistic meteorological and wave forecasts in five cases of extreme or less severe sea conditions that affected the Adriatic Sea. Some of these cases were characterized by severe damage, especially in the regions facing the northern part of the Adriatic Sea, and all of them caused flooding in the city of Venice. Two storms, those in 1966 and 1979, led to extreme flooding, while the other three of 2002, 2006 and 2008 were less severe, and caused the flooding of only a (still substantial) part of the town. All the five storms were characterized by sirocco conditions (soon to be explained). The two extreme cases of 1966 and 1979 were discussed in a recent paper by Cavaleri et al. (2010, henceforth referred to as C1), who have shown that applying the present methods and models would have provided valuable forecasts several days in advance. This result was particularly interesting because it was obtained in an enclosed basin, the Adriatic Sea, whose complicated bordering orography makes the result extremely sensitive to ever so limited modifications of the general fields. This is even truer for the marine dynamics, wind waves and surge, because of their energy dependence on the surface stress, hence on the square or more of the surface wind speeds.

While showing the potential predictability of the two extreme storms was interesting in itself, the question arises if this was the case because they were both associated with strong large-scale forcing. Indeed Grazzini (2007) has shown that severe weather events affecting northern Italy, when associated with large-scale forcing, are more predictable than normal weather conditions. On the other hand, the large quantity of, mainly satellite, data available since the early 1990s has an obvious impact on the accuracy of the daily analyses, and consequently on the quality of the forecasts. Therefore, it is of interest to explore the predictability in an enclosed sea also for relatively minor storms. This has led to the choice of the other three cases whose limited intensity, and supposedly atmospheric range, could suggest a more restricted predictability.

The first goal of this article is to document the quality of the ECMWF state-of-the-art single deterministic and ensemble-based probabilistic forecasts in predicting this type of situation. The second goal is to investigate the added value of ensemble-based probabilistic forecasts with respect to single, deterministic forecasts, and to assess their quality considering a large sample of cases.

The organisation of the paper is as follows. Section 2 describes the characteristics of the Adriatic Sea and the bordering areas. In section 3, we describe the available data and the methodology used in our simulations, and in section 4 we give a compact description of the five events we have considered. Section 5 discusses the general quality of the deterministic results and provides an initial analysis of one of the storms, aimed at finding the best way to represent the output of our research. Section 6 reports the results for all the considered storms. In section 7 we validate our results showing that the ensemble approach is not biased towards stormy conditions by discussing the EPS performance in two seasons, winters 2008–2009 and 2009–2010. Finally some conclusions are drawn in section 8.

2. The area of interest

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

The Adriatic Sea (Figure 1) is the elongated basin to the east of the Italian peninsula. It extends in the southeast to northwest direction, about 750 km long and 200 km wide. The Dinaric Alps characterise its easterly border, complemented by the Apennines on the Italian side. The bottom is deep in its central and southern parts and shallow in the north, sloping up at a rate of 1/1000 towards the northern coast. It is connected to the Mediterranean Sea via the Strait of Otranto. This limited opening and the elongated shape ensure that the wave conditions in the north depend basically only on the winds acting on the basin.

thumbnail image

Figure 1. The Adriatic Sea. The basin is about 750 km long and 200 km wide. Its easterly and westerly borders are characterised respectively by the Dinaric Alps and the Apennines ranges. The cross at its upper left end shows the position of the ISMAR oceanographic tower. Seen at its left is the lagoon where Venice is located.

Download figure to PowerPoint

All five cases discussed in this work were characterized by the presence of sirocco conditions over the Adriatic Sea, sirocco being a warm wind blowing from south to north along the sea's main axis. It is channelled in this direction by the two bordering mountain ranges. Sirocco is the main cause of the storm surges that frequently characterise the most northerly area. In front of Venice the surge may be enhanced by the sloping shallow bottom, leading to flooding of the town. C1 and Tomasin (2005) provide descriptions of the dynamics of the floods and of the factors these depend upon.

One of the difficulties of predicting weather and sea conditions in the Adriatic Sea, and generally in all enclosed basins, is the potential underestimation of the surface wind speeds predicted by numerical weather models. Cavaleri and Bertotti (1997, 2006) have repetitively shown how the ECMWF wind fields, when blowing from land to sea, are too low in the first few hundred kilometres off the coast. Given the nonlinear relationship between the energy and momentum fluxes to the sea and the wind speed, this has strong consequences on the derived marine dynamics. The level of underestimation depends on the resolution of the meteorological model. Through long-term comparisons with satellite data, both of the wind and derived wave fields, Cavaleri and Bertotti (1997, 2006) have derived a series of enhancement coefficients for the Adriatic Sea winds, typically in the range 1.2–1.5, varying in time following the progressive increase of the resolution of the meteorological model. Table I shows the coefficients used for the various resolutions.

Table I. Enhancement factors used for the ECMWF 10 m surface wind speeds in the Adriatic Sea from different resolutions of the meteorological model. The factors have been derived from extensive comparisons with satellite and buoy data.
Model resolutionEnhancement factor
T3991.42
T5111.35
T7991.26

It is clear that, while the underestimation characterises all the enclosed seas, its quantification varies from basin to basin, depending basically on its shape and the bordering orography. It is also clear that the correction should be position-dependent and varying with the meteorological pattern. However, because the overall structure of the wind fields turns out to be correct, Cavaleri and Bertotti (1997) have shown that for the Adriatic Sea a single coefficient satisfies most of the oceanographic purposes. This is particularly true for wind waves when we limit ourselves to one kind of storm, i.e. sirocco, because waves are an integrated effect, in space and time, of the driving wind fields. As such, they are not very sensitive to limited local wind differences. After many years of operational application, the significant wave height bias and scatter index (root-mean-square (rms) error divided by the mean measured value) of the model vs. measured significant wave heights (altimeters and buoys) turn out to be 0.05 m and 0.44 respectively, low enough for any practical purpose.

An alternative approach to this problem would be to use a high-resolution limited-area meteorological model. Indeed, in their analysis of the 1966 and 1979 storms, C1 have tried this approach. In that paper, all the simulations done with the ECMWF model have also been repeated with a two-level nested Bologna Limited Area Model (BOLAM) (Zampieri et al., 2005), down to a 0.06° resolution. However, granted a larger number of details of the fields, the overall result for waves and surge were not superior to the ones obtained using the enhanced ECMWF fields. It is also relevant that the ECMWF forecasts are the input information for the daily tidal forecast issued in Venice by the local Istituzione Centro Previsioni e Segnalazioni Maree (Centre for Tidal Forecasts and Warning: Canestrelli etal., 2001). Finally, in the Adriatic Sea, while there has recently been the introduction of an operational meteorological ensemble prediction system (Marsigli et al., 2008), this system is not coupled to the wave forecast. Therefore, considering also the availability of computer and data resources, for the present tests we have focused our attention only on using the ECMWF system.

3. The selected storms

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

The five sirocco storms discussed in this paper are all characterised by an intense-to-very-strong southerly flow over the Adriatic basin (Figure 1) leading to a meteorological surge in its northern part (see Table II for a list of the dates with the maximum tide). As already mentioned, the first two storms were very intense and have already been discussed in C1. The third and fourth ones were relatively less intense. The fifth one, although it was not particularly strong from the meteorological point of view, gave rise to a very severe flood in Venice, ranked fourth in history.

Table II. Date of and maximum tidal level reached during the five considered storms.
 DateMax. tidal level (m)
A4 November 19661.94
B22 December 19791.66
C16 November 20021.47
D9 December 20061.05
E1 December 20081.56

Hereafter, a compact picture of the weather conditions that led to the considered ‘acqua alta’ events in Venice is given. A still compact, but somewhat more extensive description is in appendix A to illustrate how the atmospheric flow evolved in the two days before the five events, with attention focused mainly on the low-level flow over the Adriatic Sea. Figure 2 shows the weather conditions at a time close to the peak of wind-speed intensity over the Adriatic Sea:

  • – at 1200 UTC 4 November 1966;

  • – at 0600 UTC 22 December 1979;

  • – at 1200 UTC 16 November 2002;

  • – at 0600 UTC 9 December 2006;

  • – at 1200 UTC 1 December 2008.

  • (A)
    The event of 4 November 1966—This is the historical storm with the largest-ever recorded flood. Fea et al. (1968) and Malguzzi et al. (2006), among many others, give a detailed description of the meteorological event. Tomasin (2005) and C1 provide a keen analysis of the corresponding oceanographic part. The final key meteorological factor was given by the low-level circulation centred over the Gulf of Genoa (Figure 2(a)) leading to strong surface flow along the axis of the Adriatic Sea.
  • (B)
    The event of 22 December 1979—This storm led to the second-ranked flood of Venice. The situation, shown in Figure 2(b), was in some ways similar to the 1966 case. However, in this case the meteorological system had a larger scale (Figure 2(b)) with the minimum positioned more to the west with respect to the (A) event, which yielded less extreme spatial gradients, hence wind speeds, over the Adriatic Sea.
  • (C)
    The event of 16 November 2002—This case was associated with a less severe flood of Venice. The sirocco conditions were associated with a small-scale cyclonic circulation which had developed over southern France and northwestern Italy (Figure 2(c)). When the system rapidly moved eastwards, the storm was over.
  • (D)
    The event of 9 December 2006—This case was also associated with a less severe flood of Venice. The synoptic evolution of this event has a lot of similarities with the event of 2002. A small surface low, which had developed over northwestern Italy (Figure 2(d)), led to a relatively mild southwesterly flow over the Tyrrhenian Sea and the Italian peninsula, and then was forced more along the axis of the Adriatic Sea, more so in its southern part, by the orographic effect of the Dinaric Alps.
  • (E)
    The event of 1 December 2008—This is a case of very severe flooding of Venice. The synoptic evolution of this event again has strong similarities with the previous events. Following a previous intense cyclone over France, a second small-scale cyclonic circulation developed over the Gulf of Genoa, then propagated eastwards over northern Italy. Moreover, in this case the strong westerly/southwesterly wind was channelled along the axis of the Adriatic Sea by the orographic effect of the Dinaric Alps.
thumbnail image

Figure 2. Meteorological situation at the peak of the storm of (a) 4 November 1966, (b) 22 December 1979, (c) 16 November 2002, (d) 9 December 2006 and (e) 1 December 2008. Arrow length is proportional to the local wind speed. ECMWF analysis. Surface atmospheric pressure: isobars at 4 hPa intervals.

Download figure to PowerPoint

4. Methodology

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

Forecasts for periods leading to the five storms described in the previous section have been re-generated using the state-of-the-art ECMWF global prediction system and with a high-resolution wave model of the whole Mediterranean Sea.

4.1. The ECMWF forecasting system

The dynamical part of the ECMWF Integrated Forecasting System (IFS) has been developed over a number of years jointly between ECMWF and Météo-France. The IFS model is based on the hydrostatic and shallow-atmosphere approximation of the atmospheric governing equations. It uses a horizontal spectral representation with spherical harmonics as basic functions, and a hybrid, pressure-based vertical coordinate system. The equations are integrated in time with a two-time-level semi-Lagrangian semi-implicit scheme. Physical processes include radiation, vertical diffusion, moist processes, land-surface coupling, and wave–atmosphere interaction. The reader is referred to Hortal (2004) and Beljaars etal. (2004), and references therein, for a more detailed description of the model characteristics of the systems used in this work.

ECMWF runs an operational suite that includes a high-resolution data assimilation and a single deterministic forecast (DET), and a lower-resolution ensemble prediction system (EPS). The resolution and the configuration of these three systems have been changing with time. Concerning the data assimilation (DA) and DET systems, between 2001 and 2006 they were run with spectral triangular truncation T511 and 60 vertical levels with a reduced Gaussian grid (TL511L60), moved to TL799L91 in 2006 and to TL1279L91 in January 2010. The analyses and the forecasts for the two oldest cases (1966 and 1979) were available only at a much lower resolution, i.e. TL159L31 used in the ECMWF reanalysis project (ERA-40: Uppala et al., 2005). Therefore they have been re-run at the same resolution (TL511L60) and with the same model cycle (31R1) of the 2002 case. Concerning the EPS, for the three oldest cases (1966, 1979 and 2002) the EPS has been re-run for up to 10 days with cycle 31r1 and a TL399L62 resolution, i.e. the one used operationally in 2006 and 2008. Operational TL399L62 EPS and high-resolution forecasts starting from TL799L91 analyses have been used for the last two cases. On 26 January 2010, when the resolution of the ECMWF deterministic forecast and the data assimilation were increased to TL1279L91, the EPS resolution was increased to TL639L62 up to forecast day 10, and to TL319L62 from day 10.

Mimicking ECMWF standard operational procedure, for each storm a sequence of analyses was initiated ten days before the peak of the event in the northern part of the basin. The analyses were performed with a 12-hour four-dimensional variational system, and made available at 12-hour intervals. Starting from each analysis, a 10-day DET and a 10-day EPS forecast were performed starting 10 days before each storm occurrence. The EPS initial perturbations were generated as in the operational EPS using singular vectors (Buizza and Palmer, 1995). The ensemble is made up of 50 perturbed forecasts and one unperturbed (control) forecast which uses the same model and initial conditions as the single deterministic forecast, albeit running at the resolution of the EPS. Table III shows the characteristics of the data assimilation, the deterministic model, and the probabilistic system, the EPS, used to re-run the five cases.

Table III. Resolutions and model cycles used in the data assimilation (DA), deterministic (DET) and ensemble (EPS) systems for the five storms. Txxx shows the horizontal spectral truncation with subscript ‘L’ denoting the fact that a linear reduced Gaussian grid has been used in grid-point space, and Lnn shows the number of vertical levels Model cycle 31r1 was operational at ECMWF between September 2006 and June 2007, while model cycle 35r1 was operational between January 2008 and March 2009.
StormDA analysis/DET forecastEPS
  ResolutionModel cycleResolutionModel cycle
A4 November 1966TL511L60 (∼60 km)31r1TL399L62 (∼80 km)31r1
B22 December 1979.”.”.”.”
C16 November 2002.”.”.”.”
D9 December 2006TL799L91 (∼35 km).”.”.”
E1 December 2008.”35r1.”35r1

4.2. The WAM wave model

The WAM model has been used for wave model simulations. WAM is a well-established advanced third-generation wave model amply used by the international community. It is based on the spectral description (frequency and direction) of the wave conditions at each node of the grid describing the basin. The energy balance equation, complemented with a suitable description of the relevant physical processes, is used to follow the evolution of each wave spectral component. A full description is given by Komen et al. (1994) and Janssen (2008). A six-day run, preceded by a three-day warming-up, was done for each available wind forecast, deterministic and ensemble. In addition, an analysis run was done using as input the sequence of analysis fields.

A 0.25° resolution was used for the whole Mediterranean Sea. Following the discussion in section 2, the wind speeds in the Adriatic were enhanced according to the resolution of the meteorological model (T399, T511 or T799). Janssen (2008) has provided a clear picture of the present situation and performance of the model. For the remaining Mediterranean Sea the wind speeds were corrected using the corresponding coefficients derived from the MEDATLAS study (Cavaleri and Sclavo, 2006) based on a long-term comparison between satellite and model data. While this was done for consistency, please note that, as already mentioned, the ‘out of the Adriatic’ wave data have practically no influence on the wave conditions in the northern part of this basin, where we are presently focusing our attention. This has been verified also with some direct tests. Thirty frequencies and 24 directions were used for each WAM run.

Past long-term experience at the regional tidal forecast centre (Canestrelli et al., 2001) suggests that 3 days is the limit for which operational tidal forecasts provide regular and reliable warnings. In our work we have decided to test whether valuable information could be available for a longer lead time of up to 6 days. Therefore the wave model (WAM) has been run for 6 days, using the forcing fields from the ECMWF forecasts.

5. Quality of single deterministic forecasts

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

To introduce the reader to the challenges linked to the prediction of severe sea conditions in the Adriatic Sea, this section starts with a more-detailed discussion of the storm of 2008.

Figure 3 shows the overall wind and wave fields (analysis) on the Adriatic Sea at the peak of the 2008 storm. The south/southeasterly wind is evident throughout the basin (left panel), generating (right panel) relatively large waves propagating towards north/northwest. Please remember that, apart from the possible coupling effects between waves and surge dynamics, waves may be a major factor in establishing the sea level at the coast. The loss of momentum flux associated with bottom-induced breaking in shallow water leads to an increase of the sea level towards the coast, the set-up, that in unfavourable circumstances can be one metre or more (Longuet-Higgins and Stewart, 1964; Bowen et al., 1968; Guza and Thornton, 1981). On the other hand, the sea, and in particular the enclosed seas, has a strong memory. Once out of balance, the Adriatic Sea has two dominant seiches, with periods about 11 and 22 hours (Tomasin et al., 2005). Therefore the actual sea level experienced during a storm depends also on the previous history of the basin (days) and on the astronomical tide (on the Venetian coast the spring tide excursion is about one metre). Crucial for the overall level is the phase among seiches, the astronomical tides, storm surge and wave set-up.

thumbnail image

Figure 3. Wind (a) and wave (b) fields in the Adriatic Sea at the peak of the storm of 1 December 2008. Isolines at 4 m s−1 and 1 m intervals respectively.

Download figure to PowerPoint

An ideal location for monitoring the atmospheric and oceanographic conditions in front of the Venice lagoon is the Istituto di Scienze Marine (ISMAR) oceanographic tower (see Figure 1 for its position). Operational since the early 1970s, it is located 15 km off the coast at a depth of 16 m. Capable of hosting people on board for several days, it provides a full set of meteorological and oceanographic data, both recorded on board and transmitted to land in real time. A full description of the tower, on-board systems and related results is given by Cavaleri (2000). Figure 4 shows that for the 2008 storm the hindcast analysis results (wind speed and significant wave height at the tower location) are close to the corresponding locally measured data. Figure 5 shows the corresponding scatter diagram and Table IV lists some related statistics for all the storms for which observed data are available. Although they do not show the very high quality of the similar results obtained in the ocean (e.g. Janssen, 2008), the analysed data is sufficiently good for most practical purposes. This is remarkable when remembering the crude, but effective, correction blindly applied to the ECMWF wind speeds on the Adriatic Sea (see sections 2 and 4).

thumbnail image

Figure 4. Time history of the wind speed (a) and significant wave height (b) at the ISMAR oceanographic tower during the event of 1 December 2008. See Figure 1 for its position. Continuous line shows the analysis data. Dots are the corresponding measured values at the model output times.

Download figure to PowerPoint

thumbnail image

Figure 5. Scatter plots of model (analysis, y) vs. measured (x) values during the event of December 2008. Wind speeds (a) and significant wave heights (b) at the ISMAR oceanographic tower. See Figure 1 for its position.

Download figure to PowerPoint

Table IV. Statistics of comparison at the ISMAR oceanographic tower (see Figure 1 for its position) between the hindcast (analysis) and the corresponding measured data. The period is the 5-day slot preceding and across the peak period of each storm. See Table II for their dates. Comparisons are analysis vs. measured data. SI, scatter index, is the rms error divided the mean measured value. X are measurements, Y analysis. Units for mean X and Y, and bias: wind (m s−1), waves (m).
 19661979200220062008
 windwaveswindwaveswindwaveswindwaveswindwaves
Best-fit slope1.150.700.860.920.880.9551.02
Mean X7.028.461.495.960.858.511.38
Mean Y8.035.941.425.220.718.291.44
Bias1.01−2.52−0.07−0.73−0.14−0.220.06
Correlation0.900.640.920.760.810.850.94
SI0.290.330.340.370.370.250.19

Let us for the time being consider only the single deterministic forecasts. Figure 6 shows the quality of 0–6-day forecasts of wind speed and significant wave height for all five storms as a function of the forecast range. For each storm the values have been normalized with respect to those from the corresponding final analysis. While this is a fraction of the overall available information, some general considerations are possible. First, the results indicate that the local wind speeds show a larger variability than the wave heights. On the one hand this reflects the critical position of the Adriatic Sea (its northern part in particular), surrounded on most of its sides by mountains, where a slight shift of the general meteorological pattern may substantially change the local wind conditions. Considering that waves are very sensitive to even limited differences of the driving wind fields, we derive the suggestion that the quality of the modelled wind fields is lower at the tower position, more in general in the northern Adriatic Sea, than in the remaining part of the basin. On the other hand, waves are an integrated effect, in space and time, of the driving wind fields, and although they refer to a specific position, they are a more robust parameter to represent the forecast quality. Secondly, in four out of five storms there is a tendency for the longer-range forecasts to underestimate the wave height. Overall, these results indicate that forecasts give a good indication of the intensity of the storm up to 4 days before the event.

thumbnail image

Figure 6. Each panel shows the ratios between the (wind speed and significant wave height) forecasts issued at different forecast ranges (days) and the corresponding final analysis at storm peak time. Five different storms are considered. The reference positions are the oceanographic tower (wind speed, see Figure 1 for its position) and a slightly more offshore position for significant wave height.

Download figure to PowerPoint

It is worth discussing in more detail the quality of the 36-hour forecast valid for the 1966 case. As C1 have pointed out, the storm and the associated surge were nicely forecast several days in advance. However, the 36-hour forecast, and only this one, failed to do so (according to it there was no flood at all) because of the anticipated passage of the strong front that changed the wind direction in the northern Adriatic Sea from southeast to west. This is not visible in the wind speed statistics (Figure 6, left panel) because this remained high also after the turn, but it can be detected in the wave statistics (right panel). Although, as just specified, the local wave conditions depend on the driving wind on the whole basin, a transverse and partly opposing wind still has some effect in lowering the local wave height.

6. Quality of ensemble-based probabilistic forecasts

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

We consider now the ensemble forecasts, firstly again discussing in more details the 2008 case. Since one of the key aims of the ensemble prediction system is to provide early warnings of potential severe conditions, attention will be focused on the 5–6-day forecast range. In particular, we want to assess whether ensemble-based probabilistic forecasts could be used to extend the range of useful forecasts also in an enclosed basin, in our case the Adriatic Sea.

Figure 7 shows the analysis from 26 November till 1 December, the deterministic forecast, the EPS control forecast (CON, a deterministic forecast run with the resolution of the ensemble), the median and the range of the ensemble forecasts issued at 0000 UTC 26 November, hence almost six days in advance with respect to the 1 December event. This is a peculiar sequence of events, with a sequence of three storms hitting the ISMAR tower between 28 November and 1 December. The flood, and the focus of our attention, is linked to the third storm.

thumbnail image

Figure 7. Ensemble forecasts of wind speed and wave height issued at 0000 UTC 26 November 2008 vs. the corresponding measured data. The target point is the ISMAR oceanographic tower (see Figure 1 for its position). AN is the analysis, DF and CF are the deterministic and control forecasts. The ensemble distribution is represented by the median and by the shadowed areas including respectively 25% and 75% of its 50 members.

Download figure to PowerPoint

It is immediately evident that the first two storms are very well predicted, both by the single deterministic forecast and the single EPS control forecast, and by the EPS. The interesting additional information that the EPS brings on top of the single forecasts is more evident in the third storm. Here the 132-hour single deterministic and EPS control forecasts underestimate the strength of the wind speed, and the wave height. By contrast, several EPS members do predict the storm. On the whole the EPS gives a probability of ∼25% that the significant wave height could be larger than 2 metres.

It is clear that to show within acceptable limits the corresponding results for all the forecast ranges and all the storms we need a different representation. However, before moving to these summarizing, but rather abstract, diagrams, it is worthwhile to get a more direct flavour of the performance of the EPS. To this aim we consider the wind field over a large area covering the Italian peninsula with focus on the Adriatic Sea. In particular, we compare the control forecasts of the 10 m wind speed versus the corresponding EPS forecasts for speeds in excess of 5 and 10 m s−1. This is done in the 120–150-hour forecast range, the exact range differing for each storm because for each storm we selected the time (range) when the EPS showed for the first time a valuable forecast probability.

6.1. Atmospheric forecasts over the Italian peninsula and the Adriatic Sea

In 1966 (Figure 8), the t+144h control forecast from 29 October 1200 UTC (valid for 0000 UTC 4 November) correctly predicts 10 m s−1 wind over the central part of the Adriatic Sea, but does not extend the area of 10 m s−1 wind to Venice, as detected in the analysis. Correspondingly, the EPS gives a 25–50% probability of wind in excess of 10 m s−1. The EPS probabilities are consistently higher for the 5 m s−1 threshold. Only 24 hours later, the t+120h control forecast from 30 October 1200 UTC correctly extends the area of wind in excess of 10 m s−1 to the coast, with the EPS giving a 50–75% probability (not shown). Thus, if a forecaster used the 25% probability of 10 m s−1 as a warning of potentially severe conditions affecting the Adriatic Sea and Venice, the EPS would have given a warning 24 hours earlier than the control forecast.

thumbnail image

Figure 8. 1966 case. (a) 10 m wind speed (10WS) in the analysis of 1200 UTC 4 November 1966. (b) 10WS t+144h control forecast started at 1200 UTC 29 October. EPS probability t+144h forecast of 10WS in excess of 5 m s−1 (c) and 10 m s−1 (d), started at 1200 UTC on 29th. Contour shading: for wind speed ((a) and (b)) every 5 m s−1, and for probabilities ((c) and (d)) for 5, 10, 25, 50, 75 and 100%.

Download figure to PowerPoint

A similar conclusion can be drawn for the 1979 case (Figure 9). The t+150h control forecast from 17 December 0000 UTC (valid for 0600 UTC on 22nd) completely fails to predict 10 m s−1 wind over the central and northern parts of the Adriatic, as detected in the analysis. On the contrary, the EPS gives a 25–50% probability of wind in excess of 10 m s−1. The situation repeats itself 24 hours later, with the CON forecast failing to predict wind in excess of 10 m s−1, while the EPS confirms a 25–50% probability (not shown).

thumbnail image

Figure 9. 1979 case. (a) 10 m wind-speed (10WS) in the analysis of 0600 UTC 22 December 1979. (b) 10WS t+150h control forecast started at 0000 UTC on 16th. EPS probability t+150h forecast of 10WS in excess of 5 m s−1 (c) and 10 m s−1 (d), started at 0000 UTC on 16th. Contour shading: for wind speed ((a) and (b)) every 5 m s−1, and for probabilities ((c) and (d)) for 5, 10, 25, 50, 75 and 100%.

Download figure to PowerPoint

The 2002 case (Figure 10) is the only one where the EPS does not give any earlier warning. In fact, the t+144h control forecast from 10 November 0000 UTC (valid for 0000 UTC on 16th) correctly predicts 10 m s−1 wind over the whole Adriatic Sea, as detected in the analysis, with the EPS giving a 10–25% probability (not shown). Similar forecasts are provided by both the approaches 24 hours later (not shown).

thumbnail image

Figure 10. 2002 case. (a) 10 m wind speed (10WS) in the analysis of 1200 UTC 16 November 2002. (b) 10WS t+150h control forecast started at 1200 UTC on 11th. EPS probability t+120h forecast of 10WS in excess of 5 m s−1 (c) and 10 m s−1 (d), started at 1200 UTC on 11th. Contour shading: for wind speed ((a) and (b)) every 5 m s−1, and for probabilities ((c) and (d)) for 5, 10, 25, 50, 75 and 100%.

Download figure to PowerPoint

The 2006 case is similar to the 1966 and the 1979 cases (Figure 11). The t+150h control forecast from 3 December 0000 UTC (valid for 0600 UTC on 9th) completely fails to predict 10 m s−1 wind over the central part of the Adriatic Sea and in front of Venice, as detected in the analysis, while the EPS again gives a 10–25% probability of wind in excess of this speed limit. EPS probabilities of wind in excess of 5 m s−1 are higher over the whole Adriatic Sea, with a 25–50% probability in front of Venice. Twenty-four hours later, the t+126h CON correctly predicts wind in excess of 10 m s−1 in the central part of the Adriatic Sea but not in front of Venice, while the EPS gives a 50–75% probability in the central part and a 10–25% probability in the upper part of the basin (not shown).

thumbnail image

Figure 11. 2006 case. (a) 10 m wind speed (10WS) in the analysis of 0600 UTC 9 December 2006. (b) 10WS t+150h control forecast started at 0000 UTC on 3rd. EPS probability t+150h forecast of 10WS in excess of 5 m s−1 (c) and 10 m s−1 (d), started at 0000 UTC on 3rd. Contour shading: for wind speed ((a) and (b)) every 5 m s−1, and for probabilities ((c) and (d)) for 5, 10, 25, 50, 75 and 100%.

Download figure to PowerPoint

The 2008 case is slightly different from the others, since in this case the EPS does not provide an earlier warning, but it does provide more consistent successive forecasts. For the 2008 storm, the t+150h control forecast from 25 November 0000 UTC (valid for 0600 UTC 1 December) correctly predicts 10 m s−1 wind over the whole Adriatic Sea, as detected in the analysis, with the EPS giving a 25–50% probability of wind in excess of this speed threshold over the central and southern part of the Adriatic Sea, and a 10–25% probability in front of Venice (not shown). Twenty-four hours later (Figure 12), the CON limits the area of 10 m s−1 wind speed to the central part of the Adriatic Sea. By contrast, the EPS probability of wind in excess of 10 m s−1 increases to 25–50% in front of Venice.

thumbnail image

Figure 12. 2008 case. (a) 10 m wind speed (10WS) in the analysis of 0600 UTC 1 December 2008. (b) 10WS t+126h control forecast started at 0000 UTC 26 November. EPS probability t+126h forecast of 10WS in excess of 5 m s−1 (c) and 10 m s−1 (d), started at 0000 UTC on 26th. Contour shading: for wind speed ((a) and (b)) every 5 m s−1, and for probabilities ((c) and (d)) for 5, 10, 25, 50, 75 and 100%.

Download figure to PowerPoint

The above figures provide a general perception of the information provided by the control and ensemble forecasts. However, for a more complete and exhaustive view of the overall results we need a different representation. This is the subject of the next subsection.

6.2. Wind speed and wave height forecasts over the Adriatic Sea and at or close to the ISMAR oceanographic tower

Let us now consider the wind speed and significant wave height forecasts at the ISMAR oceanographic tower (for wind) and at a slightly more offshore position (for waves); the latter yields a more representative position for the northern Adriatic Sea. In addition we want to provide statistics for the forecasts over the whole basin.

In Figure 7, we have seen how, for the 2008 case, the quality of the deterministic, control and ensemble forecasts varies with the forecast range. Focusing our attention on the peak time of the storm, we can extract some indicators representing the analysis, the DET and CON forecasts, the median and the representing percentages of the ensemble at this time. Having at our disposal a diagram, or the numerical results, such as Figure 7 for each forecast range, we can plot the results in a similar manner to what was done in Figure 6. In practice what we can see for each forecast range is the performance, at peak time, of the forecasts issued at each range. Figures 13 to 17 show the results of this analysis for the five storms. In these figures, panels (a) and (b) refer to the wind speed and the wave height at and close to, respectively, the location of the ISMAR tower.

thumbnail image

Figure 13. Storm 4 November 1966. Each panel shows the ratios between the various forecasts issued at different forecast range (days) and the corresponding final analysis at storm peak time. Deterministic forecast (_____), control forecast (---), median of ensemble forecasts (dotted) are shown. Light and heavily dotted areas include respectively 25% and 75% of the span of the overall ensemble forecasts. (a) Wind speed at the oceanographic tower (see Figure 1 for its position), (b) significant wave height at a point slightly more offshore, (c) ‘average’ ratios of the wind speeds over the Adriatic Sea, (d) ‘average’ ratios of the significant wave heights over the Adriatic Sea, (e) ‘average’ differences in wind direction over the Adriatic Sea, (f) ‘average’ differences in mean wave direction over the Adriatic Sea. See appendix B for the meaning of ‘average’.

Download figure to PowerPoint

To provide an overall judgement on the quality of the forecasts on the whole Adriatic, we have performed a vector analysis of the overall fields (Marsden, 1987; see appendix B for details). This provides an ‘average’ difference between the two compared fields, expressed as a ratio between the corresponding moduli and an angular difference between the corresponding directions. These results are shown in panels (c) and (d) of Figure 13 for the ratios (wind speed and wave height respectively), and in panels (e) and (f) for the corresponding differences in directions (degrees). In these panels, forecasts have been compared to analysis values. For directions, positive values indicate clockwise differences. In each panel the heavily and lightly shadowed areas include 25% and 75% respectively of the ensemble members.

Figure 13 shows the results relative to the 1966 case. The spread of the forecasts increases substantially beyond the day-4 forecast range (left side of the panels). Indeed, as pointed out in C1, it is already remarkable that, with the few data available at the time, the present forecasting systems would have been able to provide a valuable and solid forecast 4 days in advance. This is evident in panel (c), showing the wind speeds over the whole Adriatic Sea. Although the ensemble distribution tends to underestimate the wind speed, most likely in connection with the coarser resolution, it provides very consistent forecasts for the whole 6-day forecast range. Note in the corresponding directions (panel (e)) the kink to the right in the 36-hour forecast. This is associated with the anticipated (earlier than truth) passage of the front that turned the wind direction in the northern basin from south-southeast to west. As already discussed in section 5, this is not visible in the wave direction (panel (f)) as the heavy sea in the whole basin is only marginally affected by the transverse wind in its northern part. Indeed the local change at the tower is visible in panel (b). Knowing from C1 about the wrong 36-hour deterministic forecast (the implications for Venice flooding were substantial), we were interested in assessing whether the ensemble approach could somehow correct it. While we have not done a specific analysis of the derived floods (work is in progress along this line), the suggestion from panels (a) and (b) (wind and waves in front of Venice) is that by showing a ∼25% probability of what was then indeed the analysis the ensemble was indeed providing valuable information.

Figure 14 shows the results for the 1979 storm. Most of the comments on the predictability made for the previous storm also apply here. There are generally good deterministic, control and ensemble forecasts up to day 4, beyond which the single (DET and CON) forecast and some of the ensemble members diverge substantially from the analysis. There is an exception in the prediction of the wind at the tower (panel (a)), where the deterministic forecast is good up to day 6. Indeed this was instrumental in obtaining a good surge forecast in Venice up to day 6 as shown in C1, because of the extreme sensitivity of the local floods to the wind conditions in the most northerly part of the basin. Apart for the average ratios of the significant wave heights over the whole Adriatic Sea (Figure 14(d)), the EPS includes the verification analysis in its forecast range even when the single forecasts are rather far from the observed value, with the EPS predicting ∼25% probability of occurrence of the observed conditions. The errors in the single and EPS forecasts of the average ratios of the significant wave heights over the whole Adriatic Sea (Figure 14(d)) is probably due to the limited but significant errors in the prediction of the wind direction (Figure 14(e)). In fact, in a narrow and elongated basin such as the Adriatic Sea, a small shift of the direction from along the main axis to a slightly slanted direction or vice versa may change the fetch substantially, hence the wave conditions. Note that this affects only marginally the wave direction because its spectrum is dominated by the along-axis components.

thumbnail image

Figure 14. As Figure 13, but for the storm of 22 December 1979.

Download figure to PowerPoint

Before discussing the next three storms, two major points should be clearly stated. First of all, these storms happened in or after 2002, in times when the number of meteorological observations used to estimate the initial state of the atmosphere had already dramatically increased. As a result, the quality of the analyses for these three cases was higher. As a consequence the analysis error estimate was lower, leading to a smaller (say by ∼10%) impact on the initial perturbations of the EPS, which are set to reflect the analysis error. Secondly, although classified as storms, these three cases were relatively mild from the meteorological point of view, notwithstanding the fact that in terms of the sea conditions the storm of 2008 led to a very high storm surge and severe flooding in Venice (Table II). This second fact has a major impact on the ensemble spread. The combination of the slightly smaller initial amplitude of the EPS perturbations and less active meteorological conditions is reflected in the ensemble spread being, overall, smaller than for the storms of 1966 and 1979.

Figure 15 shows the results for the 2002 storm. Following the just-mentioned limited ensemble spread, the EPS forecasts show a strong consistency up to day 6. Despite a lower value of the wave heights at the offshore Venice point at day 4.5, the indication of a storm is present and consistent throughout the forecast range. The worst results can be detected in the prediction of the average ratios of the significant wave heights over the whole Adriatic Sea (Figure 15(d)), and partly of the corresponding wind speeds (Figure 15(c)). The practically perfect forecasts of the directions (panels (e) and (f)) suggest that the type of storm, i.e. its geometry, was well anticipated, but there was an underestimate of the storm intensity in the earlier forecasts (day 4 and beyond), most likely in the central and southern parts of the basin (the tower forecasts are not so different from the analysis). It is worth pointing out that the differences are in opposite directions for the deterministic and ensemble forecasts. This suggests again the critical role of resolution in establishing not only the details of the forecast fields at 4+ day range, but also some more fundamental structure, at least for what concerns enclosed seas.

thumbnail image

Figure 15. As Figure 13, but for the storm of 16 November 2002.

Download figure to PowerPoint

Figure 16 shows the results for the 2006 storm. Single and EPS-based forecasts show a very high quality for the whole forecast range at the tower location, but the deterministic forecast tends to overestimate the average wave height over the Adriatic Sea (Figure 16(d)). Note that these larger wave heights did not affect the corresponding results offshore of Venice (Figure 16(b)), suggesting that the differences were in the most southerly part of the basin. For this case, the ensemble provides remarkably accurate and consistent consecutive forecasts.

thumbnail image

Figure 16. As Figure 13, but for the storm of 9 December 2006.

Download figure to PowerPoint

Figure 17 shows the results for 2008. This is again a case of small spread, and very good ensemble forecasts for the whole 6-day forecast range. In the 4–6-day forecast range, the single high-resolution forecast underestimates the wave height (Figure 17(b) and (d))

thumbnail image

Figure 17. As Figure 13, but for the storm of 1 December 2008.

Download figure to PowerPoint

These results suggest that both single- and EPS-based forecasts are very good in the short forecast range, say up to forecast days 3–4, but this is not always the case for a longer forecast range. In fact, the results frequently indicate that in the 4–6-day forecast range the EPS-based forecast can provide valuable and more consistent information than single forecasts.

This basic positive result, that all the selected storms have been forecast at several days range, needs further verification. At this stage our conclusions have been drawn from the analysis of five storms. What we have shown is that, within the accuracy of this relatively small, although significant, sample, in such cases the system does produce a useful forecast. However, what still needs to be proved is that the system does not forecast a storm in the case of ‘no storm’. This is the subject of the next section.

7. The general statistics of the ensemble

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

Before moving to a summary and the discussion of the findings described in the previous sections, it is worthwhile to stop for a moment and consider the significance of the above results. Our aim has been to show that the use of the ensemble can extend the range of useful forecasts in cases of floods around the northern Adriatic Sea. To this purpose we have chosen five storms that led to the flooding of Venice and verified that indeed their predictability is extended when using the ensemble approach. Notwithstanding that in our set of five, three storms are relatively mild, unavoidably our results can in principle be biased towards stormy events. To clarify this point we need to consider also non-stormy cases. More generally we need to verify the general performance of the ensemble approach in a standard winter (storms, and in particular storm surges, are not very common events). For this, focusing on the meteorological aspect (10 m wind speed), we have carried out a statistical analysis of the quality of the EPF forecasts over three areas of increasing size: the Adriatic sea (28 points), the Mediterranean Sea (sea points with longitude between 1°W and 36°E and latitude between 30° and 46°N, 275 points) and the North Atlantic Ocean (sea points with longitude between 70° and 10°W and latitude between 25° and 65°N, 1040 points), for two seasons, winter 2009 (D08, JF09) and winter 2010 (D09, JF10). Furthermore, to assess whether the EPS performance was significantly different during extreme events, average statistics have been computed also for the 25 EPS forecasts starting each 5-day period centred on the day of each storm. Since attention in this work has been on forecasts up to 5–6 days, the evaluation has been limited to day-1 to day-7 forecasts. Forecasts have been evaluated on a regular 0.5° latitude–longitude grid against ECMWF analyses.

Three accuracy metrics have been used: the area under the Relative Operating Characteristics (ROCA), the continuous ranked probability skill score (CRPSS) and the ensemble spread–skill relationship. All these metrics (see e.g. Wilks (1995) for a general discussion of how to verify probabilistic forecasts) have been computed for 10 m wind speed EPS forecasts over the three areas mentioned above:

  • The ROCA, a measure of the capability of the EPS probabilistic forecasts to discriminate between the occurrence and the non-occurrence of events, has been computed for the probabilistic prediction of wind speed in excess of 10 m s−1. A ROCA of 1 indicates that the EPS is perfectly capable of discriminating between the occurrence and non-occurrence of the event, while a ROCA of 0.5 indicate no skill.

  • The CRPSS is the skill score computed from the continuous ranked probability score (CRPS) using a climatological forecast as reference, where the RPS is the equivalent of the root-mean-square error (rmse) in probability space. In other words, the RPS is the average distance between the EPS forecast probability distribution function and the observed (delta) probability distribution function. A CRPSS of 1 indicates a perfect forecast, while a CRPSS of 0 indicates that the EPS is as skilful as climatology.

  • The spread–skill relationship, defined by comparing the rmse of the ensemble-mean forecast and the ensemble standard deviation. In a well-calibrated ensemble, these two curves should be very close: an rmse higher (smaller) than the spread is an indication of under- (over-) dispersion.

Figure 18 shows the ROCA for the probabilistic prediction of the 10 m wind speed in excess of 10 m s−1, Figure 19 the CRPSS and Figure 20 the spread–skill relationship for the three regions and the three periods (winter 2009, winter 2010 and the 25 cases centred on the five selected events). ROCA results show that the EPS forecasts can discriminate between occurrence and non-occurrence, with ROCA decreasing with the forecast time from about 0.97 at day 1 to about 0.85 at day 7 for all regions. Scores vary seasonally, with ROCA overall higher in winter 2010. Note that over the Adriatic Sea the average ROCA computed for the 25 cases centred on the five selected events are similar to the values for the two winters. CRPSS (Figure 19) give a more integrated assessment (in the sense that they are not limited to only one threshold, 10 m s−1, as the ROCA) of the quality of the EPS forecasts. CRPSS values are positive for the whole forecast range, indicating that EPS forecasts are more skilful than climatological forecasts. CRPSS values are in general consistent with ROCA values. The comparison between the ensemble standard deviation and the error of the ensemble mean (Figure 20) indicates that the EPS spread is too low, especially in the short forecast range over the Mediterranean and the Adriatic Seas. The spread underestimation is less severe in the medium range (say after forecast day 3).

thumbnail image

Figure 18. Average area under the Relative Operating Characteristics curve (ROCA) computed for the probabilistic prediction of 10 m wind speed in excess of 10 m s−1 in winter 2009 (D08, JF09, dark, thick lines) and winter 2010 (D09, JF10) over the North Atlantic Ocean (top-left panel), the Mediterranean Sea (top-right panel) and the Adriatic Sea (bottom-left panel). The bottom-right panel shows the average area computed for 25 cases centred on the five extreme events (see text for more details) over the Adriatic Sea. EPS-based probabilistic forecasts have been verified against ECMWF analyses.

Download figure to PowerPoint

thumbnail image

Figure 19. As Figure 18 but for the average Continuous Rank Probability Skill Score (CRPSS).

Download figure to PowerPoint

thumbnail image

Figure 20. As Figure 18 but for the average ensemble spread measured by the ensemble standard deviation (dashed curves) and the root-mean-square error (rmse) of the ensemble-mean forecast (solid lines).

Download figure to PowerPoint

These results are in line with Sætra and Bidlot (2004)'s assessment of the accuracy of EPS forecasts (see their Fig. 11, which shows EPS 10 m wind-speed and wave-height probabilistic forecasts verified against observations). In their study, EPS forecasts at about 66 stations verified for the period September 1999 to March 2002 had positive Brier skill score up to forecast day 10. They concluded that EPS wind-speed probabilistic forecasts are very reliable.

The results shown in Figures 18–20 are also in agreement with routine evaluations of ECMWF EPS probabilistic forecasts of 10 m wind speed over Europe (not shown). It is worth mentioning that all these scores, based on the comparison against ECMWF analyses, are higher than scores computed against observations. In the case of 10 m wind speed forecasts over European land, for example, routine evaluation indicates that the ROCA for the t+144h probabilistic prediction of wind speed in excess of 10 m s−1 is about 0.9 when verified against analyses and 0.75 when verified against observations. This suggests that the ROCA values shown in Figure 18 would be up to 0.15 smaller if forecasts had been verified against observations.

Note that the results of Figures 13–17, focused (panels (a) and (b)) on the oceanographic tower and close by, could have suggested a different result. Indeed there the percentage of success (positive forecast) was larger than expected from the ensemble quantile. This clearly shows how a statistics limited in space and time, although potentially significant for a specific aim (in our cases the situation in front of the Venice lagoon), can lead to different conclusions when related to larger scales.

It is worth reminding the reader that the EPS is continuously improved. In particular, the EPS was changed three times in 2010:

  • on 26 January 2010 the horizontal resolution was increased to T639 (about 32 km) between forecast day 0 and 10, and T319 (about 65 km) afterwards (Miller et al., 2010);

  • on 22 June 2010 the methodology used to construct the EPS initial perturbations was changed, when perturbations computed from an Ensemble Data Assimilation system replaced the evolved singular vectors (Buizza et al., 2008);

  • on 9 November 2010 the stochastic simulation of model uncertainty was revised, the back-scatter scheme was introduced (Palmer et al., 2009) and the amplitude of the initial-time singular vectors was decreased by 50%.

These three changes led to a substantial improvement of the EPS spread–skill relationship, especially in the short forecast range. This means that EPS forecasts issued after 9 November 2010 should be more reliable and have higher skill.

8. General considerations

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

The performance of the wind speed forecasts provided by the ECMWF system, and the corresponding wave heights obtained using the WAM model, have been analysed for five sirocco storms in the Adriatic Sea. Our investigation has been limited to a 6-day forecast range. While, from the practical point of view this range covers the needs of any marine activity in an enclosed basin (a longer range may be required for instance for oceanic crossing), it is twice as long as the standard range for the warnings issued in Venice by the local devoted forecast Centre for the possible flooding of the town. Indeed, although the conditions for a useful forecast are different for wave heights and storm surge, we believe our results provide useful indications for a possible extension of the present standard tidal forecast range.

Our key findings can be summarized in the following three points:

  • 1 —Single, deterministic forecasts provide, on average, useful indications of an approaching storm up to four days before the event, and

  • 2 —This four-day range can be extended by one or two extra days if ensemble-based probabilistic forecasts are considered.

  • 3 —Ensemble-based probabilistic forecasts are more consistent in time with respect to single deterministic forecasts.

Although derived from a limited number of storms (five), we believe our results provide a realistic estimate of the forecast skill in these types of severe weather and sea conditions for the following reasons.

Firstly, to avoid being too biased towards extreme events, we have considered not only very extreme but also less severe, even mild, storms. The positive results previously obtained for the two strong cases had been possibly associated with the large mark/scale that this sort of event is expected to have in the atmosphere. However, having purposely chosen for verification three relatively minor storms, at least from the meteorological point of view, we found similar levels of predictability. Especially from the ensemble point of view, a thorough verification of our results implies an extension of the statistics to periods during which no waves, surge or flood were present. For this purpose, we have focused our attention on the driving force of the surges, i.e. the meteorological field, and more specifically the 10 m wind speed. To give a wider, and possibly more significant, perspective and also a comparison with more predictable environments (the open oceans), we have considered the Adriatic and Mediterranean Seas, and the North Atlantic Ocean. Two winter seasons have been considered, 2009 (D08–F09) and 2010 (D09–F10). Results showed that, on average, the EPS forecasts can discriminate between occurrence and non-occurrence and provide forecasts that are more skilful than climatological forecasts for the whole 6-day forecast length discussed in this work. These results are in line with earlier EPS evaluations reported by Sætra and Bidlot (2004).

Still, from the meteorological point of view it is worthwhile to point out that the three minor events we considered, (C), (D) and (E), were not associated with static situations, but, on the contrary, with meteorological systems evolving on a large scale. Therefore their predictability seems to be a rather general characteristic. A possible, or rather likely, reason is the enormous increase of meteorological data presently available for the atmospheric analysis, i.e. the initial condition the forecast starts from. In turn this implies better analyses, hence a longer range of useful forecasts. The three milder storms were all in recent times, which most likely compensated, as far as the results are concerned, for their smaller scale/mark with respect to the two extreme, but far in the past, events. On the other hand, apart from an historical interest in the two worst events in documented memory in the northern Adriatic Sea, we were interested in the present capabilities of the forecast system. In this respect our results are positive and suggest a possible extension of the standard forecast range presently used by the Venetian authorities. Although the dynamics of wind waves and storm surges have substantial differences, it is clear that the use of ensemble forecasts can extend the range of useful forecasts. Devoted work along this line of research is on the way. Furthermore, it is worth stressing that the first two conclusions are in line with published work (e.g. Buizza and Chessa, 2002; Buizza and Hollingsworth, 2002), and the third conclusion is in agreement with a recent investigation of the consistency of single- and ensemble-based forecasts (Zsoter et al., 2009).

Secondly, another reason in favour of a solid basis for our results is their consistency throughout the cases. Granted some expected variability, the results from the five considered storms, and in particular the three recent ones, are self-consistent, which of course adds to their reliability. The reliability of the results is supported also by the very limited spread shown by the ensemble forecasts of the three recent storms. While this is related to the relatively limited dynamics in the atmosphere in relation to the limited strength of the storms, the more fundamental reason is the more limited spread of the initial perturbations of the reference analysis in connection with its lower estimated errors.

While it was useful and meaningful to work on a specific basin, we believe our results can be extended, mutatis mutandis, to at least most enclosed basins. In this respect the Adriatic Sea can be considered as a rather ‘difficult’ basin, because of its elongated shape (high sensitivity to limited shift of the forcing wind fields) and the large orographic features that enclose the basin on three of its four sides. Therefore we suggest that the 4-day range of the useful deterministic forecasts, further extended, at least in a probabilistic sense, when ensemble forecasts are considered, can be considered as a minimum range of those expected to hold for most enclosed basins.

Our results suggest that the advantage of using more-consistent, ensemble-based forecasts rather than single higher-resolution ones is more evident when the specific local effects critically depend on when the storm hits. This is, for instance, the case for the floods that with regular and increasing frequency affect Venice and its lagoon. The three basic components of the actual tidal level: astronomical tides, seiches and surge, all have time-scales of the order of hours. As they have also the same order of magnitude (excursion of the order of 1 metre), the resulting sea level is critically dependent on their relative phase. For instance a 3-hour shift may imply, if at spring tide, a difference of 0.50 or more metres in the resulting maximum sea level. These are the type of situations when single forecasts might fail to provide the right indications, but ensemble-based probabilistic forecasts can provide a large enough signal (e.g. a 25% probability of occurrence) to warn that severe conditions might occur. For a town living less than one metre above mean sea level this is extremely relevant, and gaining 24–48 hours to prepare for a potential storm surge could help reduce damage substantially.

Three major changes have recently been introduced in the ECMWF EPS. Firstly, on 26 January 2010, the EPS resolution was increased to T639 (equivalent to about 30 km in grid-point space) from day 0 to day 10, and to T319 (equivalent to about 60 km) from day 10 onward. At the same time, the resolution of the global wave model coupled to the EPS was increased from 1° to 0.5°, the number of frequencies increased from 30 to 36 up to day 10 and from 25 to 30 from day 10 onward, and the number of directions from 24 to 36 up to day 10, and then from 12 to 24. As proved by previous tests, these changes are expected to further improve the EPS prediction of severe weather conditions, especially in areas affected by steep orography and land–sea contrasts. Secondly, on 22 June 2010, the EPS started using a new type of initial perturbations defined by an ensemble data assimilation system (EDA: Buizza et al., 2008) together with initial-time singular vectors, to define the EPS initial perturbations. Following this latest change, experimental tests have indicated that the EPS spread–skill relationship over the extratropics improved, especially in the early medium range. Thirdly, the simulation of model uncertainty has been improved on 9 November 2010, and the initial perturbations' amplitudes have been revised to improve the spread–skill relationship. These three latest EPS upgrades should further improve the value of ECMWF EPS probabilistic sea-state predictions documented in this work.

Acknowledgements

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

All the computations concerning the present research have been carried out at the European Centre for Medium-Range Weather Forecasts (Reading, UK). The authors gratefully acknowledge the help of the local User Support personnel, in particular Dominique Lucas. We also thank Anabel Bowen for her editorial work in improving the quality of some of the figures. We have appreciated the useful suggestions by one of the two anonymous reviewers that have led to the analysis shown in section 7.

Appendix A. Meteorological Situations of the Five Storms

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

With reference to section 3, we give here a still compact, but somewhat more detailed description of the meteorological situations that gave rise to the five considered storms in the Adriatic Sea.

  • (A)
    The event of 4 November 1966—Two days before the event, an upper-level trough started intensifying over the eastern Atlantic and started affecting the weather conditions over Spain. At the surface, at 1200 UTC 2 November a cut-off low started intensifying east of Spain, causing an intensification of the southerly flow in the lower troposphere. At 0600 UTC on 3rd, the upper-level trough positioned over Spain deepened, and at the surface the cut-off low almost reached the Gulf of Genoa. At this time the low-level flow over the Adriatic Sea started aligning with the axis of the basin, coming more from the southeast towards the northwest, and intensified. From this time, the flow over the Adriatic Sea kept intensifying, while keeping its alignment with the main axis of the basin. At 0000 UTC 4 November, the upper-level trough was still centred over Spain, while at the surface the minimum of the cyclonic circulation was position further east, over the Gulf of Genoa. At 1200 UTC on 4th, the mean-sea-level pressure reached a minimum value of 994 hPa at the centre of the cut-off low; this is the time when major flood events started affecting Tuscany and north-eastern Italy (Malguzzi et al., 2006). Over the Adriatic Sea, the wind kept amplifying till slightly before 1800 UTC, when it started decreasing and changing direction, linked to the passage of the low-level cut-off over north-eastern Italy. Figure 2(a) shows the weather surface conditions at 1200 UTC 4 November 1966, as depicted in the ECMWF ERA-40 analysis (Uppala et al., 2005), when the 10 m wind speed values over the Adriatic Sea reached about 25 m s−1. C1 report estimated maximum wind speed and significant wave height just offshore of Venice up to 28 m s−1 and 8 m, both fully consistent with the reported damage at the coast and the derived surge compared to the available measurements.
  • (B)
    The event of 22 December 1979—Two days before the event, an upper-level cyclonic circulation started developing over the western Mediterranean. This development was associated with a rapid low-level cyclogenesis. At 0600 UTC on 21st, the upper-level and the low-level flows over the Adriatic Sea intensified, and at the surface the flow started aligning with the basin's main axis. At 0000 UTC 22 December, the low-level cut-off low reached about 984 hPa and was positioned south of France, between Marseille and Corsica. Over the Adriatic Sea, the wind speed reached maximum values between 0900 and 1200 UTC on 22nd. Soon after, the cut-off low started moving inland, affecting the weather over southern France and losing its strength. At the same time, the low-level wind speed over the Adriatic Sea started decreasing. Figure 2(b) shows the weather surface conditions at 0600 UTC 22 December 1979, as depicted in the ECMWF ERA-40 analysis, when the 10 m wind speed values over the Adriatic Sea reached about 25 m s−1. As indicated by the meteorological surge, the conditions were not as extreme as in 1966, however heavy enough to cause severe damage to the ISMAR oceanographic tower located 15 km offshore the Venetian coast (see Figure 1 for its position). From the measurements, the hindcast (by C1) and the damage, the maximum wind speed and significant wave height at the tower were reported at 16.4 m s−1 and about 6 m.
  • (c)
    The event of 16 November 2002—Two days before the event, the upper-level flow was affected by a large-scale trough centred over the eastern Atlantic. The strong upper-level flow was also affecting the lower troposphere levels, with the whole Mediterranean Sea affected by a south- southwesterly flow. At 1800 UTC on 15th, the axis of the upper-level trough started changing orientation from southwest–northeast to south–north. As a consequence, over the central and eastern parts of the Mediterranean Sea both the upper-level and the low-level flow changed from southwesterly to southerly, and over the Adriatic Sea the wind started orienting along the sea's main axis. Between 0000 and 1200 UTC 16 November, small-scale low-level cyclonic circulations developed over southern France and north-western Italy. Over the Adriatic Sea, the wind kept intensifying till 1200 UTC and then started decreasing, in correspondence with the passage of the low-level, small-scale cyclonic systems over north-eastern Italy. Figure 2(c) shows the weather surface conditions at 1200 UTC 16 November 2002, as depicted in the ECMWF operational analysis, when the 10 m wind speed values over the Adriatic Sea reached about 20 m s−1. The maximum wind speed and significant wave height recorded at the tower were 14.8 m s−1 and 4.3 m respectively.
  • (D)
    The event of 9 December 2006—As in 2002, two days before the event, the upper-level flow was affected by a large-scale trough centred over the eastern Atlantic. The strong upper-level flow was also affecting the lower troposphere levels, with the whole Mediterranean Sea affected by a south-southwesterly flow. Already at 0000 UTC 8 December, the flow over the southern part of the Adriatic Sea reached ∼10 m s−1, flowing along the sea's main axis. At 1800 UTC on 8th, the upper-level flow started changing direction from south-westerly to southerly, and a low-level cut-off low started developing and intensifying north of Corsica. In correspondence with this development, the flow over the Adriatic Sea intensified. At 1200 UTC on 9th, the low-level cut-off low moved inland over northern Italy, thus affecting the flow over the northern part of the Adriatic Sea. Figure 2(d) shows the weather surface conditions at 0600 UTC 9 December 2006, as depicted in the ECMWF operational analysis, when the 10 m wind speed values over the Adriatic Sea reached about 20 m s−1. The maximum wind speed and significant wave height recorded at the tower were 16.2 m s−1 and 2.6 m respectively.
  • (E)
    The event of 1 December 2008—Two days before the event, on 29 November the upper-level flow was affected by a large-scale trough centred over France. This system was coupled, at the surface, to an intense cyclone, with strong westerly flow affecting the whole Mediterranean Sea. During the subsequent 24 hours, the trough deepened and the low-level wind intensified. At 0000 UTC on 30th, the low-level wind over the Tyrrhenian and Adriatic Seas started veering from westerly to southwesterly, and then southerly, reaching maximum intensity at the early hours of 1 December. This intensification and change of direction of the low-level wind was associated with the development of a small-scale cyclonic circulation over the gulf of Genoa at 0000 UTC, and its easterly propagation into the Po’ valley. Figure 2(e) shows the weather surface conditions at 0600 UTC 1 December 2008, as depicted in the ECMWF operational analysis, when the 10 m wind speed values over the Adriatic Sea everywhere are above 10 m s−1. The maximum wind speed and significant wave height recorded at the tower were 20.3 m s−1 and 3.2 m respectively. From 1200 UTC on 1st, the upper and low-level cyclonic circulation started losing intensity, and the intensity of the associated low-level wind in the Adriatic Sea progressively decreased.

Appendix B. Intercomparison of Two Vector Fields

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References

We want to intercompare two vector fields on the same grid, say b with respect to a (Marsden, 1987). We consider each vector as a complex number, i.e. a = [ax,ay] [RIGHTWARDS ARROW] (ax + i ay) = aexp(iΦ), with a the modulus and Φ the phase. If we have only one vector (i.e. one grid point),

  • equation image

provides the ratio of the moduli and the phase difference. The result of a point-by-point comparison of the two fields is another vector field. To summarize this result in a more compact way we can proceed as follows. Obviously

  • equation image

with a the complex conjugate of a. We consider the quantity

  • equation image

Ψ can be considered as the regression coefficient of the b field with respect to a. If we consider also the expression

  • equation image

this can be interpreted as a sort of minimum square quantity. Alternatively Ψ is the quantity that minimizes the distance between the two points [bi] and [Ψ ai] in an n-dimensional space (i = 1 to n).

The complex number

  • equation image

provides the ‘average’ ratio α and the ‘average’ phase difference β between the two fields.

References

  1. Top of page
  2. Abstract
  3. 1. Introduction
  4. 2. The area of interest
  5. 3. The selected storms
  6. 4. Methodology
  7. 5. Quality of single deterministic forecasts
  8. 6. Quality of ensemble-based probabilistic forecasts
  9. 7. The general statistics of the ensemble
  10. 8. General considerations
  11. Acknowledgements
  12. Appendix A. Meteorological Situations of the Five Storms
  13. Appendix B. Intercomparison of Two Vector Fields
  14. References
  • Beljaars ACM, Bechtold P, Kohler M, Morcrette J-J, Tompkins A, Viterbo P, Wedi N. 2004. ‘The numerics of physical parameterization.’ Pp 113134 in Proc. ECMWF Seminar on Recent developments in numerical methods for atmospheric and ocean modelling, 610 September 2004. ECMWF: Reading, UK. Available from http://www.ecmwf.int/publications/library.
  • Bowen AJ, Inman DL, Simmons VP. 1968. Wave ‘set-down’ and set-up. J. Geophys. Res. 73: 25692577.
  • Buizza R. 2008. The value of probabilistic prediction. Atmos. Sci. Lett. 9: 3642.
  • Buizza R, Chessa P. 2002. Prediction of the U.S. storm of 24–26 January 2000 with the ECMWF Ensemble Prediction System. Mon. Weather Rev. 130: 15311551.
  • Buizza R, Hollingsworth A. 2002. Storm prediction over Europe using the ECMWF Ensemble Prediction System. Meteorol. Appl. 9 289305.
  • Buizza R, Palmer TN. 1995. The singular-vector structure of the atmospheric global circulation. J. Atmos. Sci. 52: 14341456.
  • Buizza R, Leutbecher M, Isaksen L. 2008. Potential use of an ensemble of analyses in the ECMWF Ensemble Prediction System. Q. J. R. Meteorol. Soc. 134: 20512066.
  • Canestrelli P, Pastore F, Zampato L. 2001. ‘The monitoring network of the Centro Previsioni e Segnalazioni Maree (CPSM) of the Venice Municipalità.’ In Book of Abstracts of COST Action 40 Final Workshop ‘Sea level in Europe: Observation, interpretation and exploitation’, Dubrovnik, Croatia, 1921 September 2001.
  • Cavaleri L. 2000. The oceanographic tower Acqua Alta—Activity and prediction of sea states at Venice. Coast. Eng. 39: 2970.
  • Cavaleri L, Bertotti L. 1997. In search of the correct wind and wave fields in a minor basin. Mon. Weather Rev. 125: 19641975.
  • Cavaleri L, Bertotti L. 2006. The improvement of modelled wind and wave fields with increasing resolution. Ocean Eng. 33: 553565.
  • Cavaleri L, Sclavo M. 2006. The calibration of wind and wave model data in the Mediterranean Sea. Coast. Eng. 53: 613627.
  • Cavaleri L, Bertotti L, Buizza R, Buzzi A, Masato V, Umgiesser G, Zampieri M. 2010. Predictability of extreme meteo-oceanographic events in the Adriatic Sea. Q. J. R. Meteorol. Soc. 136: 400413.
  • Fea GA, Gazzola A, Cicala A. 1968. Prima documentazione generale della situazione meteorological relativa alla grande alluvione del novembre 1966. CNR-CENFAM PV 32: 215 pp.
  • Gouweleeuw BT, Thielen J, Franchello G, De Roo APJ, Buizza R. 2005. Flood forecasting using medium-range probabilistic weather prediction. Hydrol. Earth Syst. Sc. 9: 365380.
  • Grazzini F. 2007. Predictability of a large-scale flow conducive to extreme precipitation over the western Alps. Meteorol. Atmos. Phys. 95: 123138.
  • Guza RT, Thornton EB. 1981. Wave set-up on a natural beach. J. Geophys. Res. 86C5 41334137.
  • Hortal M. 2004. ‘Overview of the numerics of the ECMWF atmospheric forecast model.’ Pp 114 in Proc. ECMWF Seminar on Recent developments in numerical methods for atmospheric and ocean modelling, 610 September 2004. Available from ECMWF, Shinfield Park, Reading RG2 9AX, UK, or from http://www.ecmwf.int/publications/library.
  • Janssen PAEM. 2008. Progress in ocean wave forecasting. J. Comput. Phys. 227: 35723594.
  • Komen GJ, Cavaleri L, Donelan M, Hasselmann K, Hasselmann S, Janssen PAEM. 1994. Dynamics and Modelling of Ocean Waves. Cambridge University Press: Cambridge, UK.
  • Longuet-Higgins MS, Stewart RW. 1964. Radiation stresses in water waves: A physical discussion, with applications. Deep-Sea Res. 11: 529562.
  • Malguzzi P, Grossi G, Buzzi A, Ranzi R, Buizza R. 2006. The 1966 ‘century’ flood in Italy: A meteorological and hydrological revisitation. J. Geophys. Res. 111: D24106, DOI: 10.1029/2006JD007111.
  • Marsden RF. 1987. A comparison between geostrophic and directly measured surface winds over the northeast Pacific Ocean. Atmos.–Ocean 25: 387401.
  • Marsigli C, Montani A, Paccangnella T. 2008. A spatial verification method applied to the evaluation of high-resolution ensemble forecasts. Meteorol. Appl. 15: 125143.
  • Miller M, Buizza R, Haseler J, Hortal M, Janssen P, Untch A. 2010. Increased resolution in the ECMWF deterministic and ensemble predictionsystems.’ ECMWF Newsletter 124: 10–16. ECMWF: Reading, UK.
  • Molteni F, Buizza R, Palmer TN, Petroliagis T. 1996. The ECMWF ensemble prediction system: Methodology and validation. Q. J. R. Meteorol. Soc. 122: 73119.
  • Palmer TN, Buizza R, Leutbecher M, Hagedorn R, Jung T, Rodwell M, Vitart F, Berner J, Hagel E, Lawrence A, Pappenberger F, Park Y-Y, von Bremen L, Gilmour I. 2007. The Ensemble Prediction System: Recentand ongoing developments.’ A paper presented at the 36th Session of the ECMWF Scientific Advisory Committee. ECMWF Research Department Tech. Memo. 540. ECMWF: Reading, UK.
  • Palmer TN, Buizza R, Doblas-Reyes F, Jung T, Leutbecher M, Shutts GJ, Steinheimer M, Weisheimer A. 2009. Stochastic parametrization andmodel uncertainty.’ ECMWF Research Department Tech. Memo. 598, pp 42. ECMWF: Reading, UK.
  • Park Y-Y, Buizza R, Leutbecher M. 2008. TIGGE: Preliminary results on comparing and combining ensembles. Q. J. R. Meteorol. Soc. 134: 20292050.
  • Sætra Ø, Bidlot J-R. 2004. Potential benefits of using probabilistic forecasts for waves and marine winds based on the ECMWF Ensemble Prediction System. Weather and Forecasting 19: 673689.
  • Schaake JC, Hamill TM, Buizza R, Clark M. 2007. HEPEX: The Hydrological Ensemble Prediction Experiment. Bull. Am. Meteorol. Soc. 88: 15411547.
  • Taylor JW, Buizza R. 2003. Using weather ensemble predictions in electricity demand forecasting. Int. J. Forecast. 19: 5770.
  • Tomasin A. 2005. Forecasting the water level in Venice: Physical background and perspectives. Chapter 9, pp 7178 in Flooding and environmental challenges for Venice and its lagoon: State of knowledge, Fletcher CA, Spencer T (eds). Cambridge University Press: Cambridge, UK.
  • Tomasin A, Umgiesser G, Zampato L. 2005. On the dynamics of the Adriatic seiche. Pp 6574 in Scientific Research and Safeguarding of Venice, Campostrini P (ed). CoRiLa: Venice.
  • Uppala SM, Kållberg PW, Simmons AJ, Andrae U, Da Costa Bechtold V, Fiorino M, Gibson JK, Haseler J, Hernandez A, Kelly GA, Li X, Onogi K, Saarinen S, Sokka N, Allan RP, Andersson E, Arpe K, Balmaseda MA, Beljaars ACM, Van De Berg L, Bidlot J-R, Bormann N, Caires S, Chevallier F, Dethof A, Dragosavac M, Fisher M, Fuentes M, Hagemann S, Hólm E, Hoskins BJ, Isaksen L, Janssen PAEM, Jenne R, McNally AP, Mahfouf J-F, Morcrette J-J, Rayner NA, Saunders RW, Simon P, Sterl A, Trenberth KE, Untch A, Vasiljevic D, Viterbo P, Woollen J. 2005. The ERA-40 re-analysis. Q. J. R. Meteorol. Soc. 131: 29613012.
  • Wilks DS. 1995. Statistical Methods in the Atmospheric Sciences. Academic Press: San Diego, CA.
  • Zampieri M, Malguzzi P, Buzzi A. 2005. Sensitivity of quantitative precipitation forecasts to boundary layer parameterization: A flash flood case study in the western Mediterranean. Nat. Hazard Earth Sys. 5: 603612.
  • Zsoter E, Buizza R, Richardson D. 2009. ‘Jumpiness' of the ECMWF and Met Office EPS control and ensemble-mean forecasts. Mon. Weather Rev. 137: 38233836.