14.5: Forecasting El Niño

The importance of El Niño to global weather patterns has led to many schemes for forecasting events in the equatorial Pacific. Several generations of models have been produced, but the skill of the forecasts has not always increased. Models worked well for a few years, then failed. Failure was followed by improved models, and the cycle continued. Thus, the best models in 1991 failed to predict weak El Niños in 1993 and 1994 (Ji, Leetmaa, and Kousky, 1996). The best model of the mid-1990s failed to predict the onset of the strong El Niño of 1997-1998, although a new model developed by the National Centers for Environmental Prediction made the best forecast of the development of the event. In general, the more sophisticated the model, the better the forecasts (Kerr, 1998).

The following recounts some of the more recent work to improve the forecasts. For simplicity, I describe the technique used by the National Centers for Environmental Prediction (Ji, Behringer, and Leetmaa, 1998). But Chen et al. (1995), Latif et al. (1993), and Barnett et al. (1993), among others, have all developed useful prediction models.

Atmospheric Models

How well can we model atmospheric processes over the Pacific? To help answer the question, the World Climate Research Program’s Atmospheric Model Intercomparison Project (Gates, 1992) compared output from 30 different atmospheric numerical models for 1979 to 1988. The Variability in the Tropics: Synoptic to Intraseasonal Timescales subproject is especially important because it documents the ability of 15 atmospheric general-circulation models to simulate the observed variability in the tropical atmosphere (Slingo et al. 1995). The models included several operated by government weather forecasting centers, including the model used for day-to-day forecasts by the European Center for Medium-Range Weather Forecasts.

The first results indicate that none of the models were able to duplicate all important interseasonal variability of the tropical atmosphere on timescales of 2 to 80 days. Models with weak intraseasonal activity tended to have a weak annual cycle. Most models seemed to simulate some important aspects of the interannual variability, including El Niño. The length of the time series was, however, too short to provide conclusive results on interannual variability.

The results of the substudy imply that numerical models of the atmospheric general circulation need to be improved if they are to be used to study tropical variability and the response of the atmosphere to changes in the tropical ocean.

Oceanic Models

Our ability to understand El Niño also depends on our ability to model the oceanic circulation in the equatorial Pacific. Because the models provide the initial conditions used for the forecasts, they must be able to assimilate up-to-date measurements of the Pacific along with heat fluxes and surface winds calculated from the atmospheric models. The measurements include sea-surface winds from scatterometers and moored buoys, surface temperature from the optimal-interpolation data set (see Section 6.6), subsurface temperatures from buoys and XBTS, and sea level from altimetry and tide-gauges on islands.

Ji, Behringer, and Leetmaa (1998) at the National Centers for Environmental Prediction have modified the Geophysical Fluid Dynamics Laboratory’s Modular Ocean Model for use in the tropical Pacific (see Section 15.3 for more information about this model). It’s domain is the Pacific between 45\(^{\circ}\)S and 55\(^{\circ}\)N and between 120\(^{\circ}\)E and 70\(^{\circ}\)W. The zonal resolution is 1.5\(^{\circ}\). The meridional resolution is 1/3\(^{\circ}\) within 10\(^{\circ}\) of the equator, increasing smoothly to 1\(^{\circ}\) poleward of 20\(^{\circ}\) latitude. It has 28 vertical levels, with 18 in the upper 400 m to resolve the mixed layer and thermocline. The model is driven by mean winds from Hellerman and Rosenstein (1983), anomalies in the wind field from Florida State University, and mean heat fluxes from Oberhuber (1988). It assimilates subsurface temperature from the TAO array and XBTS, and surface temperatures from the monthly optimal-interpolation data set (Reynolds and Smith, 1994).

The output of the model is an ocean analysis, the density and current field that best fits the data used in the analysis (figures \(14.1.2\) and \(14.1.3\)). This is used to drive a coupled ocean-atmosphere model to produce forecasts.

Coupled Models

Coupled models are separate atmospheric and oceanic models that pass information through their common boundary at the sea surface, thus coupling the two calculations. The coupling can be one-way, from the atmosphere, or two-way, into and out of the ocean. In the scheme used by the NOAA National Centers for Environmental Prediction, the ocean model is the same Modular Ocean Model described above. It is coupled to a low-resolution version of the global, medium-range forecast model operated by the National Centers (Kumar, Leetmaa, and Ji, 1994). Anomalies of wind stress, heat, and fresh-water fluxes calculated from the atmospheric model are added to the mean annual values of the fluxes, and the sums are used to drive the ocean model. Sea-surface temperature calculated from the ocean model is used to drive the atmospheric model from 15\(^{\circ}\)N to 15\(^{\circ}\)S.

As computer power decreases in cost, models are becoming ever more complex. The trend is to global coupled models able to include other coupled ocean-atmosphere systems in addition to ENSO. I return to the problem in Section 15.6 where I describe global coupled models.

Statistical Models

Statistical models are based on an analysis of weather patterns in the Pacific using data going back many decades. The basic idea is that if weather patterns today are similar to patterns at some time in the past, then today’s patterns will evolve as they did at that past time. For example, if winds and temperatures in the tropical Pacific today are similar to wind and temperatures just before the 1976 El Niño, then we might expect a similar El Niño to start in the near future.

Forecasts

In general, the coupled ocean-atmosphere models produce forecasts that are no better than the statistical forecasts (Jan van Oldenborgh, 2005). The forecasts include not only events in the Pacific but also the global consequences of El Niño. The forecasts are judged two ways:

1. Using the correlation between the area-averaged sea-surface-temperature anomalies from the model and the observed temperature anomalies in the eastern equatorial Pacific. The area is usually from 170\(^{\circ}\)W to 120\(^{\circ}\)W between 5\(^{\circ}\)S and 5\(^{\circ}\)N. Useful forecasts have correlations exceeding 0.6.
2. Using the root-mean-square difference between the observed and predicted sea-surface temperature in the same area.

The forecasts of the very strong 1997 El Niño have been carefully studied. Jan van Oldenborg et al (2005) and Barnston et al (1999) found no models successfully forecast the earliest onset of the El Niño in late 1996 and early 1997. The first formal announcements of the El Niño were made in May 1997. Nor did any model forecast the large temperature anomalies observed in the eastern equatorial Pacific until the area had already warmed. There was no clear distinction between the accuracy of the dynamical or statistical forecasts.