ATMOSPHERIC SCIENCE LETTERS Atmos. Sci. Let. 6: 112 117 (2005) Published online 21 April 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/asl.99 Monthly forecast and the Summer 2003 heat wave over Europe: a case study Frédéric Vitart* European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, UK *Correspondence to: Frédéric Vitart, European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading, RG2 9AX, UK. E-mail: nec@ecmwf.int Received: 15 February 2005 Accepted: 18 February 2005 Abstract A monthly forecasting system has been set up at ECMWF. Probabilistic scores indicate some potentially useful skill for the periods days 12 to 18 and 19 to 32. A case study of the 2003 heat wave over Europe suggests that the model successfully predicted a risk of heat wave higher than in climatology more than 10 days in advance. Copyright 2005 Royal Meteorological Society Keywords: monthly forecast; heat wave; predictability 1. Introduction: why monthly forecasting? An operational monthly forecasting system has been set up at the European Centre for Medium-Range Weather Forecasts (ECMWF), which fills the gap between the medium-range and seasonal forecasts. It produces forecasts for the time range 10 to 32 days once a week; this range is probably still short enough that the atmosphere retains some memory of its initial state and it may be long enough that the ocean variability has an impact on the atmospheric circulation. Consequently, the monthly forecasting system has been built as a combination of the ECMWF medium-range Ensemble Prediction System (EPS) (Buizza, 1997) and the ECMWF seasonal forecasting system (Anderson et al., 2003). It contains features of both systems and, in particular, is based on coupled ocean-atmosphere integrations, as is the seasonal forecasting system. The 10 to 32 days time range is often considered difficult for weather forecasting. However, an important source of predictability in the monthly time range over Europe in the 10- to 30-day range is believed to originate from the Madden Julian Oscillation (MJO) (see, for instance, Ferranti et al., 1990). Several article (see, for instance, Flatau et al., 1997) suggest that oceanatmosphere coupling has a significant impact upon the propagation of MJO events in the Indian Ocean and Western Pacific. The use of a coupled system may, therefore, help capture some aspects of MJO variability. A thorough assessment of the overall skill of the ECMWF monthly forecasting system is discussed in Vitart (2004) over a smaller sample (45 cases instead of 76 in the present study), and without discussing any individual case. 2. Description of the monthly forecasting system The monthly forecasts are based on an ensemble of 51 coupled ocean-atmosphere integrations (one control and 50 perturbed forecasts). The length of the coupled integration is 32 days, and the frequency of the monthly forecasts is currently once a week, which is more frequent than the seasonal forecasts (once a month), but less frequent than the EPS (twice a day). The atmospheric component is the ECMWF atmospheric model Integrated Forecast System (IFS), using the same model version as the deterministic forecast. Currently, the atmospheric model is run at T L 159 resolution (1.125 1.125 ) with 40 levels in the vertical. The oceanic component is the Hamburg Ocean Primitive Equation (HOPE) model (Wolff et al., 1997), from the Max Plank Institute for Meteorology, Hamburg. The ocean model has lower resolution in the Extratropics, but higher resolution in the equatorial region, in order to resolve ocean baroclinic waves and processes, which are tightly trapped at the equator. The ocean model has 29 levels in the vertical. The atmosphere and ocean communicate with each other through the Ocean Atmosphere Sea Ice Soil (OASIS) coupler (Terray et al., 1995). The atmospheric fluxes of momentum, heat and fresh water are passed to the ocean every hour and, in exchange, the ocean sea surface temperature (SST) is passed to the atmosphere. The frequency of coupling (every hour) is higher than in seasonal forecasting (every 24 h), since high frequency coupling may have some impact on the development of some synoptic scale systems, such as tropical cyclones. A more complete description of the ECMWF monthly forecasting system can be found in Vitart (2004). After 10 days of coupled integrations, the model drift begins to be significant. The effect of the drift on the model calculations is estimated from previous integrations of the model in previous years (the back-statistics). The drift is removed from the model solution during the post-processing. In the present system, the climatology (back-statistics) is taken from a 5-member ensemble of 32-day coupled integrations, Copyright 2005 Royal Meteorological Society
Monthly forecast and the summer 2003 heat wave over Europe 113 starting on the same day and month as the real-time forecast for each of the past 12 years. This represents a 60-member ensemble. 3. Product and potential predictability Monthly forecast products include anomaly, probability and tercile maps based on comparing the 51- member ensemble distribution of the real-time forecast to the 60-member ensemble distribution of the model climatology (the back-statistics). The forecasts are based on weekly means. Fields like surface temperature, 2-m temperature, precipitation and mean sealevel pressure have been averaged over 7 days. The 7-day periods correspond to days 5 to 11, 12 to 18, 19 to 25 and 26 to 32. These periods have been chosen so that they correspond to Monday to Sunday calendar weeks (the monthly forecast starting date is on Thursday at 00 UTC). Anomalies (relative to the model climatology from the past 12 years) are derived. Figure 1 displays an example of anomaly maps of geopotential at 500 hpa over the Northern Extratropics. (a) A Wilcoxon Mann Whitney test (WMW test, see, for instance, Wonacott and Wonacott, 1977) is applied to estimate whether the ensemble distribution of the real-time forecast is significantly different from the ensemble distribution of the back-statistics. Regions where the WMW test displays a significance, less than 90% are blank. In other words, shaded areas in the plots represent areas where the model displays some potential predictability. Unsurprisingly, in the vast majority of cases, the percentage of areas of potential predictability decreases week by week, indicating that the model is drifting towards its climatology. In general, the model displays strong potential predictability over a large portion of the Extratropics for the period days 12 to 18. However, there is generally a sharp decrease of potential predictability in the last two weeks of the forecasts. After 20 days of forecasts, the ensemble distribution is generally close to the model climatology, and it is not rare that the model displays very little potential predictability over Europe at this time range. However, there have been some cases where the monthly forecasting system displayed strong potential (b) (c) (d) Figure 1. Weekly mean ensemble mean anomalies of 500 hpa geopotential height relative to the model climate predicted by the monthly forecasting system starting on 31 December 2003. Blue corresponds to negative anomalies and red corresponds to positive anomalies. The fields are displayed using a contour with intervals of 2 dam starting at 2 dam (red colour) and 2 dam (blue colour)
114 F. Vitart predictability in the last two weekly periods over the Northern Extratropics (North of 30 N). The forecast starting on 31 December 2003 (Figure 1) is one of those cases, and, in fact, displayed the strongest potential predictability of all the forecasts since March 2002, when the monthly forecasting started to be run routinely. Interestingly, the forecasts for days 19 to 25 and 26 to 32 are clearly different from the forecast of days 12 to 18 (Figure 1b). Therefore, this strong potential predictability in days 19 to 25 and 26 to 32 is not due to the persistence of the initial conditions or the medium-range forecasts. This forecast verified quite well. Most especially, the low pressure over Europe predicted in the two last weeks of the forecast was observed, although its centre was slightly more to the east in the verification (not shown). As a consequence, the monthly forecast predicted cold conditions over Europe more than 20 days in advance, as observed. This forecast, which displayed the strongest potential predictability since the start of the monthly forecasting runs in March 2002, displayed the highest anomaly correlation score for days 19 to 25 and 26 to 32 over the Northern Extratropics. This suggests that the model may display some reliability after 20 days. However, this was just an example, and the next section will discuss the general reliability of the system over a large number of cases. 4. A brief assessment of performance After 10 days, the spread of the ensemble forecast starts to be large, and the forecasts are essentially probabilistic. Probabilistic scores of the monthly forecasting system have been evaluated through the scores obtained with weekly averaged surface temperature, 2- m temperature, precipitation and mean sea-level pressure. Seventy-six cases have been verified, covering all seasons since March 2002 (start of the ECMWF monthly forecasting system). The relative operating characteristics (ROC) scores are based on contingency tables giving the number of observed occurrences and non-occurrences of an event as a function of the forecast occurrence and non-occurrence of that event (see Stanski et al., 1989 for example). The monthly forecasting system displays ROC scores based on weekly means that are generally higher than climatology during the period days 12 to 18. Figure 2a displays the ROC diagram of the probability that the 2-m temperature is in the upper tercile for days 12 to 18 over all the land points in the Northern Extratropics (North of 30 N) calculated over all the 76 cases. The ROC area is of order of 0.7, which is well above the climatological value of 0.5. A grid-point map suggests that the model performs better than persistence at this time range over the vast majority of the grid points. This is true for all the seasons, but most especially in winter. However, for this time range, a tougher test consists in comparing Figure 2. ROC diagrams of the probability that 2-metre temperature averaged over a) days 12 18 and b) days 19 32 is in the upper tercile. The red curve represents the diagram obtained with the monthly forecasting system, and the blue curve represents the diagram obtained by persisting the probabilities from the previous weekly periods ( a) days 5 11 and b) days 5 18). The ROC diagrams have been computed over 76 cases covering all seasons the skill of the monthly forecasting system to the skill of the forecast obtained by persisting the probabilities of the previous weekly period (days 5 11). For days 12 to 18, the model performs significantly better than persistence of days 5 to 11 probabilities (blue line in Figure 2). A scatter plot diagram shows that the monthly forecasting system beats persistence in more than two-third of all the cases, and the difference is statistically significant according to the WMW test. Figure 3a displays the reliability diagram (Wilks, 1995) of the probability that the 2-m temperature anomaly averaged over days 12 to 18 exceeds 2 K for all the land points in the Northern Extratropics. This graph displays the observed frequency as a function of the forecast probability. Figure 3a shows that the model displays some reliability, with the observed frequency increasing with the forecast probability. Similar results to those above are attained for all the other variables that have been verified, such as mean sea-level pressure, precipitation, surface temperature and for different thresholds of the probabilistic event. However, the scores display some variability.
Monthly forecast and the summer 2003 heat wave over Europe 115 but still remain generally higher than climatology and persistence (Figure 2b). However, at this time range, the value of the forecast depends essentially on the threshold of the probabilistic event. The model displays much more useful skill and reliability in predicting large anomalies like the probability that the 2-m temperature anomaly averaged over the period days 19 to 32 is larger than 2 K (Figure 3b). 5. A case study: heat wave over France in August 2003 Figure 3. Reliability diagram of the probability that the 2-metre temperature anomally averaged over the period a) days 12 18 and b) days 19 32 exceeds 2K in the Northern Hemisphere extratropics. Only land points have been are considered. For the model, the anomalies are relative to the model climate, for observations they are relative to ERA40 or ECMWF operational analysis. The size of the circles on the reliability curve is proportional to the population of the forecst probability bin. The dotted diagonal represents the diagram of a perfect forecast. The dashed horizontal line represents the diagram of a no-skill forecast. The dashed vertical line represents the model climatology For instance, probabilistic scores obtained with precipitation are significantly lower than those with the surface temperature, but the main conclusion remains that the monthly forecasting system performs significantly better than climatology and the persistence of days 5 to 11 probabilities. For the period days 19 to 32, the scores are unsurprisingly much lower than in days 12 to 18, The heat wave during the first two weeks of August 2003 over Europe was particularly intense. The overall impact on society has been exceptional, with severe disruption of activities and heavy loss of life in many European countries. Health authorities estimated that, because of the soaring temperatures, about 14 000 died in France alone, and thousands more casualties were reported in other countries. The present study focuses on the week from 3 to 9 August, when 2-m temperature anomalies relative to the past 12-year climate were close to 10 K over some parts of France (Figure 4a). Figure 4b displays the monthly forecasts starting on 30 July 2003, and shows the ensemble mean of 2-m temperature anomaly (relative to the model climatology) for the period day 5 to 11. At that time range, the model predicted a strong positive anomaly of the 2-m temperature, although the intensity of the ensemble mean is less than in the analysis (Figure 4b). Almost all the ensemble members predicted a significant heat wave, but only a few members of the ensemble predicted an intensity as strong as in the analysis. At that time, the monthly forecasting system was running every two weeks, but, for this test case, a hindcast starting on 25 July 2003 has been produced to evaluate the weekly evolution of the forecast. A subjective test suggests that 15 members of the ensemble predict a weekly 2-m temperature anomaly larger than 3 K over most of Europe. Four ensemble members display an anomaly with amplitude of the same order as in the analysis. Figure 5 displays the 2-m temperature anomaly over Europe predicted by one of these ensemble members. In the 60-memberensemble model climatology corresponding to this real-time forecast, only four members display a weekly 2-m temperature anomaly larger than 3 K, and none of them has an amplitude comparable to the analysis. This suggests that although most ensemble members do not predict a significant heat wave over Europe for the period days 12 to 18, the model predicted a probability for this extreme event higher than in the model climatology. The ensemble mean (Figure 4c) displays a significant and positive anomaly in 2-m temperature averaged over days 12 to 18 over most of Europe. However, since the anomaly is averaged over all the ensemble members, its magnitude is less than in the analysis.
116 F. Vitart (a) (b) (c) (d) Figure 4. The top left panel represents the 2-metre temperature anomalies from the ECMWF operational analysis (relative to the past 12-year climate computed from the ERA-40 and operational analyses) averaged over the period 3 9 August (the verification). The top right panel represents the 51-member ensemble mean of the 2m temperature anomalies from the monthly forecasts starting on 30 July 2003 (the 2-metre temperature anomalies have been averaged over days 5 11 and they are relative to the model climatology). Only areas where there is a significant difference between the ensemble distribution of the real-time forecase and the model climatology according to the WMW- test are shown coloured. The bottom left panel represents the 2m temperature anomalies from the hindcast starting on 25 July 2003, and averaged over all the ensemble members and the period days 12 18. Finally, the bottom right panel shows the 2-metre temperature anomaly from the monthly forecast starting on 16 July 2003, and averaged over all the ensemble members and the period days 19 25 Figure 5. The 2m temperature anomalies averaged over the period 3 9 August. (a) The 2-metre temperature anomaly from the ECMWF operational analysis. (b) The 2-metre temperature anomaly predicted by one member of the ensemble (ensemble number 2) and starting on 25 July 2003. The contour interval is 1K. Areas where the anomaly exceeds 2K are coloured. This figure indicates that some members of the ensemble predicted a significant heat wave over Europe more than ten days in advance Maps of individual ensemble members for the period days 19 to 25 of the monthly forecast starting on the 16 July 2003 indicate that 12 ensemble members predict a weekly 2-m temperature anomaly larger than 3 K, with one forecast displaying a positive 2-m temperature anomaly as strong as in the analysis. In the model climatology corresponding to this forecast, only 4 ensemble members display a weekly 2-m temperature anomaly larger than 3 K, and none of them has an amplitude comparable to the analysis. This suggests that 20 days in advance, the model predicted a higher probability than usual of a significant heat wave over western Europe. The ensemble mean (Figure 4d) displays a strong (although weaker than in the analysis) positive anomaly over a large portion of Europe. Such a strong signal in the ensemble mean is quite unusual for this time range. 6. Conclusion A monthly forecasting system has been set up at ECMWF. The model performs generally better than
Monthly forecast and the summer 2003 heat wave over Europe 117 persistence of the probabilities from the previous week and better than climatology. For days 19 to 32, the model displays more skill in predicting large anomalies. This suggests that the model can be useful at the time range days 10 to 30 and is likely to be more useful than a monthly forecast using persistence of medium-range weather forecasts and climatology. Therefore, extending dynamical forecasts beyond days 10 seems useful in the intra-seasonal time range. The case study of the 2003 heat wave over Europe indicates that the model did not predict in a deterministic way that the first week of August will have an exceptional 2-m temperature anomaly over western Europe. It is not clear if this is due to a lack of predictability of this type of event or if this is due to a shortcoming of the monthly forecasting system. Nevertheless, the model was successful in predicting a higher probability of this extreme event happening. The monthly forecasting system could be useful by providing an early warning of this type of extreme event. References Anderson D, Stockdale T, Balmaseda M, Ferranti L, Vitart F, Doblas- Reyes P, Hagedorn R, Jung T, Vidard A, Troccoli A, Palmer T. 2003. Comparison of the ECMWF seasonal forecast systems 1 and 2, including the relative performance for the 1997 1998 El-Nino. ECMWF Technical Memorandum 404. http://www.ecmwf.int/publications/library/do/references/list/14. Buizza R. 1997. Potential forecast skill of ensemble prediction and spread and skill distributions of the ECMWF ensemble prediction system. Monthly Weather Review 125: 99 119. Ferranti L, Palmer TN, Molteni F, Klinker E. 1990. Tropicalextratropical interaction associated with the 30-60-day oscillations and its impact on medium and extended range prediction. Journal of Atmospheric Science 47: 2177 2199. Flatau M, Flatau PJ, Phoebus P, Niiler P. 1997. The feedback between equatorial convection and evaporative processes: the implication for intraseasonal oscillations. Journal of Atmospheric Science 54: 2373 2386. Stanski HR, Wilson LJ, Burrows WR. 1989. Survey of common verification methods in meteorology. World Weather Watch Technical Report No. 8, WMO tech. DOC. 358, 114. Terray LE, Sevault E, Guillardi E, Thual O. 1995. The OASIS coupler using guide version 2.0. CERFACS, Technical Report No. TR/CMGC/95-46, Toulouse, France, 123. Vitart F. 2004. Monthly Forecasting at ECMWF. Monthly Weather Review 132(12): 2761 2779. Wilks DS. 1995. Statistical Methods in the Atmospheric Sciences: An Introduction. Academic Press: San Diego; 464. Wolff JO, Maier-Raimer E, Legutke S. 1997. The Hamburg ocean primitive equation model. Deutsches Klimarechenzentrum Technical Report No. 13, Hamburg, Germany, 98. Wonacott TH, Wonacott RJ. 1977. Introductory Statistics. John Wiley: New York; 650.