Progress towards reliable and useful seasonal and interannual climate predictions

s towards reliable and useful seasonal and interannual climate predictions

by Tim Palmer*

Based on a scientific lecture given to the WMO Executive Council in 2004, this article describes progress in the field of seasonal climate prediction. Through the World Climate Research Programme (WCRP), WMO has played a crucial role in the development of seasonal climate prediction. For example, in the Tropical Ocean Global Atmosphere (TOGA) project, the basic paradigm was established that tropical air-sea interactions, particularly those associated with El Niño, can impart seasonal predictability onto remote parts of the global atmosphere. The Climate Variability and Predictability project, CLIVAR, which followed TOGA, brought seasonal climate prediction to a state where operational predictions based on global coupled ocean-atmosphere models were possible. Applications projects such as WMO’s Climate Information and Prediction Service (CLIPS) are now exploring the potential of seasonal climate prediction for benefiting society as a whole.

Introduction

The transition from research to operations and to potential applications requires that seasonal forecasts are reliable. As discussed below, this means that we must have a clear understanding of the uncertainties involved in seasonal predictions, and are able to express these uncertainties in the form of forecast “error bars”. In this way, the user can know how certain the forecasts are likely to be. One of the key uncertainties in climate prediction is associated with ill-defined procedures in the computational representation of the basic equations of motion of climate. In this paper, the design of reliable seasonal forecast systems using multi-model ensembles, and their corresponding application in health, agronomy and water-management activities is discussed.

What is predictability?

A mid-life celebration of CLIVAR’s achievements was recently held in a major conference in Baltimore, USA,  attracting over 600 participants. CLIVAR’s mission statement is:

To observe simulate and predict Earth’s climate system, with a focus on ocean-atmosphere interactions, enabling better understanding of climate variability predictability and change, to the benefit of society and the environment in which we live.

As this statement clearly enounces, a goal of CLIVAR is an improved understanding of the predictability of climate. A further goal is an enablement of such an understanding to benefit society in general. Are these sequential goals? Should we try to apply CLIVAR science only when we can first fully quantify the predictability of climate? The purpose of this section is to suggest that the study of predictability is not an “ivory-tower” pursuit, but something that is intrinsically linked with the applications for which the forecasts are made. The two goals are, in some sense, parallel and complementary goals.

To pursue this, let us ask: what exactly do we mean by predictability? Consider some meteorological variable, for sake of illustration: summer mean temperature over Geneva. The year-to-year variation of this variable can be described in terms of a climatological probability distribution, represented schematically as the solid curve in Figure 1(a) as a normal or Gaussian distribution.

Figure 1 — (a) The solid line is a schematic illustration of the climatological probability of some climatic variable, such as seasonal-mean surface temperature for Geneva. The dashed line is a schematic illustration of the seasonal forecast probability distribution showing clear predictability; (b) as (a) but for a forecast probability distribution whose overall level of predictability is small. However, for certain applications which require knowledge of whether or not temperature exceeds some threshold, there may be significant predictability.

Now, suppose we have a forecast system that predicts the probability distribution of summer mean temperature for Geneva, a season ahead. Such a system is necessarily based on the ensemble forecast methodology, discussed below. Suppose that such a system predicts, for a particular summer, the dashed curve in Figure 1(a). The two probability distributions are clearly different, indicating unambiguous predictability for this summer’s prediction: that the seasonal-mean temperature will be above average.

What about the situation in Figure 1(b)? Throughout most of the temperature range, the difference between the forecast probability distribution and the climatological probability distribution is small. As such, we might be inclined to say that, in this situation, predictability is, in general, rather small. But suppose we are mainly interested in the prevalence of a certain weather-sensitive disease, call it “X”, which only becomes prevalent if the temperature exceeds some threshold Tc, or suppose that some crop“Y” fails if temperature exceeds Tc. Then for these health and agricultural applications, this particular forecast probability distribution would imply very useful and important predictability: that the probability of disease X becoming prevalent or of crop Y failing, is predicted to be very small in the coming season.

These two specific examples (X and Y) are simplistic; however, the point is that assessment of whether predictability exists or not is inherently linked to the applications for which the seasonal forecasts are being used. Indeed, this motivates a plausible definition of predictability: a variable x is predictable if  the forecast probability distribution of x differs sufficiently from the climatological probability distribution to influence relevant decision-makers.

Some more realistic examples of this applications-driven approach to the issue of quantifying climate predictability are given below. However, a few words need to be said first about how forecast probability distributions can be made. Let us start by considering weather prediction.

Ensemble weather prediction

As we all know, the initial conditions for a weather forecast are never known perfectly. However, with modern-day supercomputers, we can run weather forecast models many times from very slightly different initial conditions, consistent with the uncertainties—at ECMWF we run the medium-range forecast model 52 times twice a day. The resulting forecasts can be combined to produce a forecast probability distribution.

Figure 2 is an example of a 42-hour ensemble forecast for the infamous and highly-damaging storm Lothar of December 1999. ECMWF’s main high-resolution deterministic forecast completely missed this storm. In this particular case, the spread of the ECMWF ensemble was enormous, indicating that the development of the flow was highly unpredictable. However, despite this unpredictability, the ensemble did indicate a significant probability, or risk, of a severe weather event. Of course, 42-hour weather forecasts are not normally this unpredictable.

Figure 2 — Surface pressure “stamp maps” for the 42-hour ECMWF ensemble forecast for the severe storm Lothar

The scientific basis for ensemble forecasting can be demonstrated using Lorenz’s original mathematical model for chaos, based on a simplification of the atmosphere’s equations of motion, as shown in Figure 3. The evolution of three different ensembles are shown. The point is that, because the underlying equations are non-linear, the growth of initial uncertainty is strongly dependent on the starting conditions. The practical point of this is that the predictability of forecasts is variable; we need ensembles to tell us in advance how predictable the climate system is.

Figure 3 — The scientific basis for ensemble forecasting illustrated using the prototypical Lorenz (1963) model of low-order chaos, showing that in a nonl-inear system predictability is flow dependent.

There are many ways to demonstrate the skill of an ensemble forecast. However, ensemble forecasts are also economically valuable for predicting weather-sensitive risk (Palmer, 2002). A convincing way of showing the value of ensemble weather forecasts over the conventional deterministic forecasts is through the type of analysis illustrated in Figure 4, obtained through a collaboration between ECMWF and the London School of Economics. We imagine a casino taking bets on the temperature at London, Heathrow in the coming days. A gambler bets on certain temperatures in proportion to their probability of occurrence as predicted by the ECMWF ensemble prediction system. By contrast, the casino pays out according to the prediction made by the ECMWF high-resolution deterministic forecast. To prevent the house from going bankrupt in short time, the high-resolution forecast is dressed by applying a Gaussian probability distribution based on its average past error. Figure 4 shows that the gambler will make money against the house in both the early and late medium range—more so for the latter than the former.

Figure 4 — An assessment of the economic value of the ECMWF Ensemble Prediction System, in comparison with the deterministic forecast (L. Smith and M. Roulston, personal communication)

Before returning to the problem of seasonal climate prediction, I want to discuss the problem of how ensemble prediction can be used in the broadcast media. In fact, ensemble prediction is already being used in some European countries. Often the claim is made that there is not enough on-air time to talk about probabilities, and that the public may not understand these probabilities anyway. My own feeling about this issue is that this issue can be resolved by making better use of the World Wide Web. On-air forecasters should remark that their forecasts are the most likely scenarios, but, for perfectly good scientific reasons (cf Figure 3), predictions are necessarily uncertain, sometimes more uncertain than other times. The forecasters can then refer to some online site reference where probabilistic predictions can then be shown, at least for key cities, e.g. as in the example in Figure 5 (for Geneva).

Figure 5 — An example of a routine medium-range ECMWF “EPSgram” for Geneva. It is definitely going to get colder and drier for the coming weekend, though, interestingly, temperatures are more uncertain for Thursday, 19 August than they are for Saturday, 21 August.

Ensemble prediction for seasonal climate prediction

Let us now return to seasonal climate prediction. The physical basis for seasonal climate prediction lies in components of climate that vary slowly compared with individual weather events, i.e. ocean and land surface (including cryospheric components). As is well known (thanks to projects such as TOGA and CLIVAR), El Niño is the prototypical phenomenon with predictability on the seasonal time-scale. In order to predict seasonal climate by dynamical means, fully coupled ocean-land-atmosphere models are required. Just as in weather prediction, ensemble forecasts using these coupled models give probabilistic risk forecasts of climate events. However, for seasonal ensemble prediction it is essential to take into account not only uncertainty in the initial conditions, but also uncertainty in the model equations themselves. This latter uncertainty arises because the process of parametrization, the way in which sub-gridscale motions are represented in weather and climate models, is not a precisely defined procedure.

One way to represent model uncertainty is to incorporate within the ensemble, completely different models. The resulting ensemble prediction system is known as a multi-model system. (There are other ways, e.g. stochastic physics (Palmer, 2001)).

The skill and utility of multi-model ensemble forecast systems has been explored in a recent European Union project called DEMETER (Palmer, 2004). Demeter was the Greek goddess of fertility. When her daughter Persephone was abducted into the underworld, Demeter cast a deep freeze on the Earth, so that no crops would grow. Eventually Zeus intervened and Persephone was allowed out for nine months of the year. Demeter kept the freeze on the Earth for the remaining three months. Without Demeter we would not have seasons, and hence seasonal forecast projects! More prosaically, DEMETER stands for Development of a European Multi-Model System for Seasonal to Interannual Prediction. In this project, advantage was taken of one of the real strengths of European climate research: the existence of quasi-independent state-of-the-art comprehensive global climate models developed at a number of institutes around Europe.

Figure 6 gives an example of how the DEMETER multimodel forecast system produces more reliable seasonal forecasts. Shown are two- to four-month forecasts of El Niño sea-surface temperature anomalies over the period 1980-2001. The red dot shows the observed El Niño sea-surface temperature anomaly, the green “bar and whiskers” show the forecast probability distribution of El Niño sea-surface temperature anomaly in terms of terciles. Forecasts at the top of Figure 6 are based on the ECMWF model only. Whilst there is clear skill, the ensemble system is not fully reliable: there are many cases where the verification lies outside the range of the ensemble. The bottom diagram is for the full DEMETER multi-model ensemble system; now the verification almost always lies with the range of the ensemble. This is one example of many that show conclusively that the DEMETER ensemble is intrinsically more useful and more skilful than forecasts from any one (e.g. national) model. If ever there was a reason for international collaboration on climate prediction, this is it!

.
Figure 6 — Top: seasonal forecast probability distributions of El Niño, based on the ECMWF model within the DEMETER project. The red dot gives the observed value, which often lies outside the forecast. Bottom: as top, but for the DEMETER multi-model ensemble. The validation now almost always lies within the ensemble. The probabilistic skill of the DEMETER multi-model ensemble is greater than for the single-model ensemble.

A key part of the DEMETER project was to demonstrate value for applications in health and agriculture. In order to study this quantitatively, a system was developed using partners expert in downscaling, and partners expert in malaria prediction, on the one hand, and crop modelling on the other. Figure 7 is an example of the month 2-4 DEMETER forecast probability distribution of malaria prevalence for a grid point in southern Africa (Morse et al., 2005). As with the El Niño sea-surface temperature distribution, terciles of the forecast probability distribution are shown. There is some clear year-to-year variability, and the “verification” (here based on running the malaria model using ERA-40 weather) lies within the forecast ensemble.

Figure 7 — Forecast probability distributions of malaria prevalence over southern Africa, based on the DEMETER multi-model ensemble system coupled to a malaria prediction model. The blue dot denotes the ensemble mean value, and the red dot denotes malaria prevalence when the malaria model is forced with ERA-40 gridded analyses (so-called tier-2 validation. (Source: Morse et al., 2005)

Studies of malaria prediction for Botswana, using DEMETER data, has been made by colleagues at the Columbia University Earth Institute (Thomson et al., 2004). The authors of this study were so impressed with results from the DEMETER forecast system that they claimed that it was now possible to realize the aspiration of the Abuja Declaration of African Health ministers, to provide reliable forecasts of malaria prevalence ahead of the start of the rainy seasons.

Figure 8 shows month 2-4 forecast probability distributions from DEMETER, but here based on wheat yield in tonnes/hectare over specific European countries (Canteloube and Terres, 2005; see also Marletto et al., 2005). Whilst only a limited number of years have so far been studied, there is evidence of useful predictability, and the European Commission’s Agriculture Directorate-General has expressed considerable interest in these products for their operational crop yield assessments.

Figure 8 — Probability distributions of wheat yield for different countries in Europe, based on DEMETER multi-model re-forecasts (from Canteloube and Terres, 2005). The blue dot denotes the ensemble mean value, and the red dot denotes crop yield when the crop model is forced with ERA-40 gridded analyses. The bars denote the Eurostat yield.

DEMETER data are also of value for crop yield prediction in the tropics (groundnuts in Gujurat; Challinor et al., 2005). Also, in collaboration with the Georgia Institute of Technology and the Bangladesh Meteorological Institute, the skill of the DEMETER forecasts coupled to hydrological models for the Bangladesh drainage basins, are being used to forecast probability distributions of flooding in Bangladesh (P.J. Webster and T. Hopson, personal communication).

Scientists who wish to assess the extent to which there is useful predictability for their part of the world for particular applications of interest are strongly encouraged, to go to the DEMETER Website. All the DEMETER data can be freely downloaded from this Website, as can the ERA-40 data used for validations.

Results from the DEMETER project are based on “re-forecasts” over the ERA-40 period of re-analyses. A legacy of DEMETER, however, is a real-time seasonal forecast system, currently based on ECMWF, Met Office and Météo-France climate models. An example of a pair of real-time products from this real-time system is shown in Figure 9.

		Figure 9 — Example  of a real-time multi-model seasonal forecasts, based on the Met Office, UK and ECMWF seasonal forecast models.

Throughout this paper, the inherent value of seasonal forecasts for a range of applications has been discussed. However, seasonal prediction also plays a role in improving numerical weather forecasting, and for validating multi-model ensemble prediction used for climate change, e.g. in the IPCC AR4 process. There is not space in this paper to discuss these important aspects of seasonal prediction work.

Conclusions

In conclusion, recent work to recognize explicitly the inherent uncertainty in the formulation of climate models has led to a reliable seasonal ensemble forecast system, which can be used in health, agriculture, hydrology, water management and a range of other applications. Research leading to such a multi-model system exploits the sort of international collaboration promoted by WMO. The DEMETER project will be further developed in the EU FP6 project ENSEMBLES, which in turn will contribute significantly to WCRP’s new COPES initiative. The multi-model system is now being assessed within numerical weather prediction as part of the THORPEX project.

References

Canteloube P. and J.-M. Terres, 2005: Use of seasonal weather forecasts in crop yield modelling. To appear in Tellus 57A (also available from: http://www.ecmwf.int/research/demeter/ news/tellusa.html)

Challinor, A.J., J.M. Ó, T.R.Wheeler and F.J. Doblas-Reyes. Probabilistic simulations of crop yield over western India using the DEMETER seasonal hindcast ensembles. To appear in Tellus 57A (also available from http://www.ecmwf.int/ research/demeter/news/tellusa.html)

Hagedorn, R., F.J. Doblas-Reyes and T.N. Palmer, 2004: The rationale behind the success of multi-model ensembles in seasonal  forecasting. To appear in Tellus 57A (also available from: http://www.ecmwf.int/research/demeter/news/tellusa. html)

Lorenz, E.N.. 1963. Deterministic non-periodic flow. J. Atmos.Sci. 42: 433-471.

Marletto, F. Zinoni, L. Criscuolo, G. Fontana, S. Marchesi, A. Morgillo, M.R.M. Van Soetendael, E. Ceotto and U.  Anderson, 2005. Evaluation of downscaled DEMETER multi-model ensemble seasonal hindcasts in Northern Italy by means of a model of wheat growth and soil water balance. To appear in Tellus 57A (also available from http://www. ecmwf.int/research/demeter/news/tellusa.html).

Morse, A.P., F.J. Doblas-Reyes, M.B. Hoshen, R. Hagedorn, M.C. Thomson and T.N. Palmer, 2005: First steps towards the integration of a dynamic malaria model within a probabilistic multi-model forecast system. To appear in Tellus 57A (also available from http://www.ecmwf.int/research/demeter/ news/tellusa.html)

Palmer, T.N., 2001: A nonlinear dynamical perspective on model error: a proposal for nonlocal stochastic-dynamic parametrisation in weather and climate prediction models. Q. J. R. Meteorol. Soc., 127, 685-708.

Palmer, T.N., 2002: The economic value of ensemble forecasts as a tool for risk assessment: from days to decades. Q. J. R. Meteorol. Soc., 128, 747-774.

Palmer, T.N., A. Alessandri, U. Andersen, P. Canteloube, M. Davey, P. Délécluse, M. Dequé, E. Díez, F.J. Doblas-Reyes, H. Feddersen, R. Graham. S. Gualdi, J.-F. Guérémy, R. Hagedorn, M. Hoshen, N. Keenlyside, M. Latif, A. Lazar, E. Maisonnave, V. Marletto, A.P. Morse, B. Orfila, P. Rogel, J.-M. Terres and M.C. Thomson, 2004: Development of a European multi-model ensemble system for seasonal to inter-annual prediction. Bull. Amer. Meteor. Soc., 85, 853-872.

Thomson, M.C., S.J.Mason, S.J.Connor, T.Phindela, F.J.Doblas-reyes, R.Hagedorn, A.P.Morse and T.N.Palmer, 2004: Malaria epidemics predicted using DEMETER seasonal climate forecasts. Submitted to Science.

_________

*European Centre for Medium-Range Weather Forecasts (ECMWF), Shinfield Park, Reading, RG2 9AX, United Kingdom