«Medium-Range Weather Prediction Austin Woods Medium-Range Weather Prediction The European Approach The story of the European Centre for Medium-Range ...»
After making the first forecast, small changes were made to the analysis, changes that were within the range of the known errors of the analysis. Thus, the second analysis was in principle almost as accurate as the first, but it was different, and a forecast run from this would differ from the first. This would be repeated several times, so that an ensemble of different forecasts, all for the same time, would be made. Since each forecast had in principle similar accuracy, the ensemble could be examined statistically for the likelihood — or the statistical probability — of precipitation, cold or hot spells, strong winds and more.
A substantial research effort devoted to the assessment of predictability on the monthly and seasonal timescales was now starting. It was based on chaos theory, one of the major scientific developments of the twentieth century. Chaotic systems are governed by precise equations that determine their evolution, but they are characterised by behaviour that is unpredictable and seemingly random. The equations are said to be “non-linear” and are unstable to small perturbations. The EPS provided a practical tool for estimating how small differences in the analysis could affect the subsequent forecast.
Thus, Prof Ed Lorenz’s concept of chaos theory was to be applied with a practical goal. Lorenz developed his theory to study the range of predictability of the atmosphere, an inherently chaotic system. At the Centre, many numerical “butterflies’ wings” were to be flapped in the model’s atmosphere; the resulting different forecasts would be examined statistically to determine the predictability of the real atmosphere.
There were preconditions for a successful outcome. The model used should have no large systematic errors; the results would be only as good as the model. The size of the ensemble should be large; small samples would produce unreliable statistics and probabilities. Theory suggested that an ensemble of about 50 members would be required to account for the different structures possible. Very powerful computers would be required.
Starting with 24 members in initial experiments, the sizes of the ensembles had reached 32 by 1992. Since running the ECMWF operational model to ten days took about two hours, the ensemble had to be run at a lower resolution, such that an individual ten-day prediction would be completed in Ensemble prediction — forecasting the error 119 about two minutes. Clearly, more computer power would allow the resolution of the model, and the size of the ensemble, to be increased in the future.
The EPS approach could in principle be applied to the model as well as to the analysis; for example, the parameterisation of the small-scale properties in the model could be perturbed. Thus, to take into account the effect of uncertainties in the model formulation, each forecast can be made using slightly different model equations. Work continued over the following years, and by winter 1992–93, a real-time EPS experiment was under way.
From December 1992, the ECMWF operational medium-range numerical prediction system was made up of two elements. One was the operational forecast produced using a model with 31 levels capable of resolving atmospheric waves with a resolution down to 190 km. After almost nine years of experimentation in the field, and at first only three times a week, the other element was an EPS using a model at the lower resolution of 700 km and 19 levels. In this first system, only the analyses were changed. The uncertainties arising from model errors were not taken into account. At this time NCEP too started to produce operational EPS forecasts.
The Centre’s pioneering Ensemble Prediction System started to provide a growing range of new products to help forecasters deal scientifically and quantitatively with the day-to-day variations in the predictability of the atmosphere. The EPS allows forecasters to predict the skill of the operational forecast objectively — to forecast the forecast skill.
In July 1993, participants from ten Member States attended a two-day Expert Meeting on the EPS at the Centre. They reviewed the status of the still experimental system. How large should the dispersion of the forecasts in the EPS be? Too small, and the different forecasts lie closer to each other than to the verifying analysis; too wide, and the statistics would not be useful. It was clear that the most important EPS products would be probabilities of temperatures being significantly above or below normal, the so-called “anomaly”, and precipitation.
Making realistic initial perturbations turned out to be a key factor, and an interesting scientific challenge. Early attempts essentially added random noise at each grid point. This did not work. The model by and large simply dissipated the resulting perturbations into the flow. Instead, it was necessary to change or perturb the analyses in unstable regions, and to perturb them in the right way.
Information on the inherent dynamical instabilities of the flow was used.
The perturbations had to be designed to represent the uncertainties of the operational analysis. The “spread” of the forecasts in an ensemble could be increased or decreased simply by increasing or decreasing the amplitude of
120 Chapter 10the perturbations applied to the analysis. However, in practise it soon became clear that increasing the amplitude resulted in an increasing number of poor forecasts. A correlation was found between the skill of the forecasts and the amount that the ensemble spread, a necessary precondition for the viability of the EPS.
From May 1994, the EPS was run daily, instead of three days a week. The value of the EPS for predicting occurrences of severe weather, strong winds or heavy rain for example, as well as its use for prediction of forecast skill, was recognised by the SAC in early 1994. For example even if say only 5 to 10 forecasts in an ensemble of 50 were to predict an unusually severe storm a week from now, this would be taken as a first warning of an event to be monitored with care in later forecasts.
An evaluation of the EPS in mid-1996 showed that the system provided “non-trivial” information about the forecasts out to the limit of the Centre’s operational prediction, to ten days ahead. Probabilities of temperature anomalies showed a significant degree of skill. Two problems were recognised.
Although reduced, systematic model errors can never be eliminated, especially with the rather low resolution required to run a large number of forecasts. In addition, there was insufficient spread in the EPS. More powerful computing could alleviate these, allowing increased resolution and more forecast runs, i.e. a larger sample size.
A major upgrade to the EPS was introduced in December 1996: now there were 50 members instead of 32, and the resolution of the model was increased to 31 levels with the grid spacing reduced to 120 km. We note that the Centre was now running 50 forecasts each day at the resolution of the operational medium-range model five years earlier!
In late 1996, a study using a high-resolution T213 31-level EPS system showed how the system could be used to give a measure of confidence in forecasts of extreme rainfall during intense Mediterranean storms. Three cases were studied: in all three, the high-resolution prediction indicated extreme precipitation. In two cases, one over Italy and the other over Greece, the EPS suggested a high probability of such precipitation, and heavy rain did occur. In the Italian case, which occurred in November 1994, catastrophic flooding and land slides over northern Italy and southern France led to the loss of more than 60 lives. Over Greece in October 1994, heavy rainfall in the region around Athens caused the loss of 12 lives and much property damage. In the third case, the EPS gave a low probability, thus not supporting the high-resolution forecast of intense precipitation over northern Italy. This was correctly identified by the EPS as a false alarm.
The Centre’s operational prediction of the severe floods over Europe in January 1995 was consistently successful. This was due in large part to the Ensemble prediction — forecasting the error 121 good performance of the EPS, which gave consistently high probabilities of heavy precipitation.
An investigation into the possibility of using models or analyses from other forecast systems, e.g. that of the UK Met Office, was made; however available evidence indicated that it was the analysis differences that were important, rather than model differences, in producing the required divergence in the forecasts making up the ensemble. In 1995, EPS results were being exchanged between the Centre, the UK Met Office and NMC Washington, and performances of the differing systems were being compared.
The EPS produced huge amounts of data: 50 different forecasts to ten days ahead of all weather parameters for the entire globe. How can such veritable avalanches of data, produced daily, be best presented to a potential user? First, of course, the probability distribution of any weather parameter anywhere can be determined. We have seen that probabilities of temperature anomalies and rainfall can usefully be derived. Beyond this, “clustering” and “tubing” of forecasts were investigated. “Clusters” of several forecasts in the ensemble brought together those forecasts that were on the whole similar. For example, 10 of the 50 forecasts with a predominantly northerly flow over Europe might form one cluster, 7 or 8 with mainly anticyclonic flow another and so on. For a “bench forecaster” who has to make up his mind how to present the weather for the week ahead on TV, such clusters, stressing similar forecasts in the ensemble, were useful tools.
“Tubing” of the forecasts took a different approach. It could be assumed that the ensemble mean is more likely to be the best indicator of the future weather. “Tubes” of the different forecast elements were derived, leading from the central group of forecasts to the different extremes. Thus, forecasts in the different tubes all differed in a similar way from the mean.
Both clustering and tubing were designed to facilitate an interpretation by the human forecaster of the large volume of EPS information, and complemented well the probability information.
In 1998, the EPS model was again enhanced. Uncertainties that the analysis system had detected were added to the uncertainties growing rapidly at the beginning of the forecast. Now also, the system was taking into account model uncertainties caused by known errors in the model’s treatment of physical processes in the atmosphere. The scheme to do this, known as “stochastic physics”, had been developed and implemented by Miller, Palmer and Buizza; it introduced a random noise into the equations. Many advantages resulted from the changes of 1996 and 1998: the ensemble mean was more skilful, the spread of the predictions was improved, and the probabilities became more reliable.
122 Chapter 10In 1998, David Richardson carried out some work at the Centre to address the question: what is the economic value of the EPS? Is it in fact worth the cost? If on being given a forecast, a user decides to take action that he would not otherwise have taken, and benefits economically from this, the forecast will have been of value to the user. Indeed before an offshore oil rig costing hundreds of millions of Euros to build can be towed from its port of manufacture to its eventual site, the operator must be able to convince his insurer that he has obtained the best weather forecast for the route and for the duration of the tow, a period of perhaps several days. A full analysis of the benefits of a forecast system requires detailed knowledge of the weather-sensitivity of the application, and the decision-making process of the user.
Richardson examined the case of a decision-maker who can choose to take action or do nothing, and the resulting cost and/or loss. For example, the cost could be to “grit the roads”, and the loss would be that arising if frost occurred and the roads remained without grit. The advantage of EPS probabilities became evident; the user can select a probability threshold appropriate to his needs. Richardson showed that a six-day EPS forecast at the then level of accuracy would provide about 60% of the savings that would be gained with a perfect knowledge of the future weather.
In November 2000, with more powerful computers, the EPS was again enhanced: the resolution was now increased to 80 km. The vertical resolution had been increased to 40 levels the previous October. The pace of change was accelerating. Now each of the 50 forecasts run daily had a higher resolution than that of the main medium-range model in use at the beginning of 1998. The performance of new system was compared to that of the old. As would be expected, there was a significant gain in predictability, of about 12 hours in fact. The higher resolution EPS was generally better able to predict the intensity of severe storms, even to about six or seven days ahead. In particular, experiments showed the EPS to be better capable of predicting the intensity and the position of the severe storms that affected Europe in December 1999.
It was now evident that the EPS had reached a mature stage. Its output products were suitable for use in weather risk management. The storm in France in December 1999 caused about €10 billion damage. Weather-related damage increased in frequency during the 1990s. Demand for relevant information increased from commercial interests as well as from the public. It was increasingly recognised that a single forecast can fail to indicate the intensity, location or timing of a severe weather event. A study by Roberto Buizza in 2001 showed how the EPS could be used to update and Ensemble prediction — forecasting the error 123 refine a-priori estimates of possible losses, and to quantify the probability that a “maximum acceptable loss” will occur. The work was extended to reduce errors in predicted energy demand using EPS predictions of wind, cloud cover and temperature.
Frederico Grazzini and Francois Lalaurette developed two new tools to help condense the massive flow of information from the EPS system. The “EPS-gram” summarises the time sequence of weather at a single point. The “Extreme Forecast Index” identifies the likely occurrence of significant but rare weather events.
The increased accuracy of the EPS predictions was quantified in 2001.