degree. Indeed, these forecasts are usually more accurate than the information
inferred for a given location from scattered weather stations at other locations.
Paradoxically, though, if the models that generate the forecasts aren’t
regularly updated with real-world data, the forecasts will rapidly become no better
than the predictions from the Farmer’s Almanac. That’s because weather is
chaotic (in both a technical and nontechnical sense). Left undisturbed, the output
of even the best model will get further and further away from the real weather
over time.
So the best approach is to use a bit of each: Rely on the forecasts when
they’re reliable, turn to the weather station data when they’re not. The problem of
combining the two effectively is called data assimilation, and solving it requires
some pretty fancy mathematical tricks.
The challenge is to figure out when the forecasts are likely to be right and
when they’re questionable. When a strong cold front pushes through the eastern
U.S. in September, for example, several days of clear weather will almost
certainly follow. In that case, the forecast is likely to be better than the
measurements. But when a hurricane is churning northward along the East
Coast, its path is hard to predict, so the forecasts can’t be relied on and real-
world measurements are essential. Mathematics is needed to distinguish
between these situations.
One of the best approaches to this at the moment is called a “Local
Ensemble Transform Kalman Filter” (LETKF). A LETKF creates a collection of
forecasts, not just one: Researchers run the model fifty times using slightly
varying initial data. If those simulations lead to fifty very different results in one
area, the researchers know that the forecast is highly uncertain there, and they
rely heavily on the data from the weather stations. But in regions that come out
pretty much the same, they trust the model’s forecast more than the data. In
areas with few weather stations (like the middle of an ocean), the LETKF is as
much as 65% better than the techniques currently in use. But even better
techniques are still needed.
Data assimilation was initially developed in the context of weather
forecasting, but it could be used in areas as diverse as oil recovery, CAT scans,
forestry, fisheries, or maybe even climate. Any situation in which a model makes