|
Predictive Models
Statistical & Dynamical Scientists build
theoretical models for several reasons. One of these is to try to understand the
underlying cause of a phenomenon. Biologists, for instance, build models of how animals
and plants grow to better understand the growth processes which are taking place. The
second reason to build models is to predict future events. For instance, if a model can
simulate ocean temperature and height in a previous El Niņo episode, it may also be able
to simulate these same parameters in future El Niņos.
El Niņo models take ocean parameters [wind speed,
ocean temperature, atmospheric pressure, rainfall, sea height data, etc.] for the past few
months. The models predict what will happen in the future. The model data will hopefully
answer the following questions. Will an El Niņo develop? Will it be large or small? Will
there be a following La Nina episode, and how large will that be?
There are two main types of forecasts.
First there are statistical forecasts, based on historical records. Second are dynamical
forecasts, based of forward integration of numerical models of the coupled
ocean-atmosphere system. Each has its strengths and weaknesses, and the results from these
can be quite different.
| Statistical Models |
Statistical forecasts correlate observed weather conditions with occurrences
of El Niņo. Typically, sea surface interactions (SST) in the key regions of the
equatorial Pacific are used to define "El Niņo periods". Alternatively an
index known as the "Southern Oscillation Index" (SOI) is used, based on the
surface pressure difference between Tahiti and Darwin. The advantage of the SOI over SST
is that the SOI records go back at least a century, while we have only a few decades of
SST observations in mid-ocean. Then the correlation of one of these indices with, for
example, rainfall in California, is the basis for a forecast of the likelihood of
reoccurrence of heavy rains in that region during an El Niņo winter. These are probably
the most common type of forecast that is seen on the media. In some regions, such as the
US Gulf Coast, these correlations are quite robust and the statistical forecast is fairly
reliable. In others the correlations are weak and/or marginal.
The strength of statistical forecasts is that they
are based on events that actually did occur. However, they can fail because El Niņo is
not an exact, repeating phenomenon. We observe that different events evolve in different
patterns, can occur at different times of the year, and so on. In addition, there are many
climate oscillations occurring simultaneously, and the present weather at any location is
the sum of these oscillations and the interactions between them. Therefore, it is
not straightforward to isolate the specific effects of El Niņo by averaging over previous
events. All these things result in blurring the statistics and reducing the confidence in
such a forecast.
Another problem with statistical forecasts is that
we do not have good, long-term records of many of the important quantities of interest.
Once you go back further than the mid-1950s, the ocean records are sparse and ambiguous,
making it hard to determine which are strong El Niņo years and which are weak ones.
However, if attention is limited to the period of "good" data, then there are
really only a handful of events, and the statistics become quite unreliable. Many of
the differences among statistical forecasts reported in the media are due to the choice of
different averaging periods. |
The bottom picture is an acutal image from a
satellite that shows how much the computer modeled image [top] underestimated this El
Niņo. Shades of reds show the degree of sea level above average.
Thanks to NASA/JPL/Caltech
| Dynamical Models |
Dynamical forecasts are based on hydrodynamical equations numerically
integrated forward from present observed conditions. These computer models range from
relatively simple representations to complex models such as are used in weather
forecasting. During the 1980s it appeared as if El Niņo could be explained by planetary
waves bouncing around the Pacific, and this could be depicted easily in a computer model.
However, this theory failed to predict the events of 1990s, proving to us that we must
incorporate the full complexity of the ocean-atmospheric system in the simulation. This is
a task of utmost difficulty since it compounds the problems of ordinary weather
forecasting by the addition of numerous interactions between the ocean and the atmosphere.
A major difficulty in this type of forecasting is
that we cannot simulate every molecule of air and water. Thus, at many times, these
simulations turn out be crude, blunt grid mesh representations of the earth. Furthermore,
due to computer speed and storage, these grids have spacing of typically tens to hundreds
of kilometers. Take, for example, the representation of clouds in such models. The grid is
far too coarse to resolve individual clouds, and therefore, many clouds are combined to
act as a whole. To correctly predict the amount of water and heat released by a
could, we have to know the actual speed and humidity of rising air. Thus, the amount of
precipitation produced by a group of individual clouds is not the same as that which would
be produced by a cloud that had the average properties of the whole region. Much
current research is devoted to figuring out how to represent complex interactions like
these in a way that computers can work with.
Nevertheless, as techonlogically-inclined students,
it is our belief that as computer become faster and as our understanding of the physical
processes becomes better, we will rely more and more on the dynamical forecasts.
They have the tremendous advantage of working forward from the actual present observed
conditions, and so avoid the problem of statistically averaging over a number of events
that differ in important details. In addition, for low-frequency events like El Niņo, it
will take decades or centuries to accumulate sufficient realizations to really improve
statistical confidence. This maybe so because this field offers the opportunity for
scientists to make significant progresses by advancing the understanding of physical
processes within the coupled system, as we have already seen over the past several years. |
|
