'Superensemble' technique combines strongest elements of several predictions
By FLORENCE OLSEN
Everyone knows weather forecasts are wrong a fair
amount of the time. Now one of academe's most respected climate researchers is
turning some of those mistakes into a resource to improve the predictions of
computerized weather models.
The researcher is T.N. Krishnamurti --
"Krish" to his friends here at Florida State University. Mr. Krishnamurti, a
professor of meteorology, is the creator of one of the world's two dozen or so
computer programs for modeling and predicting weather and climate change around
the world.
More recently, he has become a champion of a new
weather-prediction technique known as superensemble forecasting, in which a
computer program essentially merges the predictions of different computerized
weather models, taking from each the elements on which it has the best track
record. That's where Mr. Krishnamurti's database of proven mistakes comes in --
he uses it to evaluate each model's strengths and weaknesses. He hopes the
superensemble technique will lead to more objective and reliable long-range
forecasts.
It was a superensemble forecast that earned Mr. Krishnamurti
his 15 minutes of network-television fame last year: His technique correctly
predicted the path of Hurricane Floyd four days before the storm made landfall
near Wilmington, N.C. Forecasters at the National Hurricane Center in Miami had
predicted that the storm would move farther west -- an erroneous forecast that
triggered a massive evacuation from the coastal communities of Florida, Georgia,
and South Carolina.
This hurricane season has been less eventful for Mr.
Krishnamurti. As he clicks to open up a view of his own virtual-climate system,
he describes the immense data feeds that generate the bright yellow, orange, and
dark-red contours displayed on his computer screen. "We're taking data from the
Arctic to the Antarctic, and from the surface of the Earth all the way up to the
stratosphere and above," Mr. Krishnamurti says, stretching his arms toward the
ceiling of his office. "We take everything," including data from five U.S.
military satellites.
Global models like Mr. Krishnamurti's are complex
software programs made up of mathematical equations representing the laws of
atmospheric physics -- insofar as they are known. To complicate matters, natural
feedback processes give weather its bedeviling trait of "nonlinearity": The
weather continually reacts to itself, with unpredictable results. Computer
models make their predictions by solving nonlinear equations of fluid dynamics.
The three-dimensional computer models are among the few tools that
scientists have for studying how the world's climate is changing and getting
warmer. Although models have detected the warming trend, scientists say the
models haven't answered the more controversial question of whether human
activities are the cause. Only a limited number of global climate models exist
throughout the world -- some say 20, others say 30 or more -- and no two are
alike. But all of the models require, Mr. Krishnamurti says, "an immense amount
of computing."
Ever since the first computer modeling experiments of the
1950's, meteorologists have used the ever-greater capacity of computers to add
more realism, or resolution, to their models. "There's a lot of pride involved
in getting a model to represent accurately what's going on," says Christian D.
Kummerow, an associate professor of atmospheric sciences at Colorado State
University.
Before his meteorology lab began acquiring multiprocessor
computers about five years ago, Mr. Krishnamurti and his graduate students ran
large-scale weather and climate simulations on supercomputers owned by the
Energy and Defense Departments and the National Center for Atmospheric Research,
a research facility in Boulder, Colo., operated by a consortium of North
American universities. Mr. Krishnamurti's picture hangs on a wall at NCAR. "He's
always been their No. 1 customer," says C. Eric Williford, a Ph.D. candidate in
meteorology at Florida State who has been a coauthor of some of Mr.
Krishnamurti's papers.
So it comes as no surprise to his Florida State
colleagues that Mr. Krishnamurti will be the first researcher to run a
large-scale model on the university's new $8-million supercomputer. When it is
fully installed next year, the terascale computer -- an International Business
Machines RS/6000 SP2 with 680 processors -- will have a peak capacity of 2.5
trillion calculations per second. In addition to his position in the meteorology
department, Mr. Krishnamurti is program director for computational climate
dynamics in the university's new School of Computational Science and Information
Technology, whose faculty members and graduate students have full access to the
supercomputer.
Mr. Krishnamurti's passion is his climate model, which he
has spent more than two decades perfecting and which he calls the Florida State
University Global Spectral Model. Over the years, his graduate students have
contributed 45 or more dissertations to the collective work.
For climate
researchers like Mr. Krishnamurti, the modeling problem can be maddening and
beautiful at the same time. The researchers readily acknowledge numerous ways
that computer models can go wrong in making both long- and short-range
forecasts. The models themselves may make faulty assumptions about the processes
that cause the weather and climate to change. The data fed into the models is
often erroneous. The models and the data may not be completely tuned for
efficient computation. "We are dealing in the world of error," Mr. Williford
says.
"In theory, if you had a perfect model with infinite resolution,
and you had infinitely good initial data, then you'd get a perfect forecast,"
says John P. Boyd, a professor of atmospheric, oceanic, and space sciences at
the University of Michigan at Ann Arbor. In reality, Mr. Boyd says, weather and
climate models rely on equations to describe the relationships between what are
merely average values for atmospheric properties within imaginary "grid boxes"
superimposed over the real atmosphere. "There are necessarily errors made in
doing this," he says.
Indeed, weather and climate are far too complex
for any one model to get every forecast right. But some models are better than
others, depending on the forecast locations and conditions, Mr. Krishnamurti
says. Much of his research has involved painstakingly identifying the systematic
forecast errors produced by his own and other global models, then storing those
in a huge "errors database" that has grown to 10 million statistics.
Mr.
Krishnamurti collects these statistics to improve his ability to make forecasts
using models. To assess how realistic their hurricane forecasts are, for
instance, Mr. Krishnamurti's research group measures the differences between
their model's forecast of a particular hurricane's path and the actual path of
that hurricane. "That's hindcasting," he says.
Recent efforts to produce
more accurate model forecasts have led meteorologists to devise a variety of new
forecasting techniques, and ensemble forecasting is prominent among them. It
"has a lot of potential for doing good," Michigan's Mr. Boyd says. The technique
involves running the same forecast model, say, 10 different times, starting each
time with slightly different data about the initial weather conditions. In a
variation on the technique, researchers run 10 different forecast models with
the same initial data to measure the extent to which the forecasts agree.
"If all the forecasts are very similar, it means you can trust that
forecast more," says Zoltan Toth, leader of the global prediction group of the
General Sciences Corporation at the National Weather Service's Environmental
Modeling Center, in Camp Springs, Md. But if the forecasts differ a great deal,
it means "on that particular day you cannot give a really precise forecast."
A year ago, Mr. Krishnamurti incorporated into his research what he
describes as an enhanced ensemble technique. Mr. Krishnamurti calls it
"multimodel superensemble" forecasting. His method, which differs from the way
other meteorologists do ensemble forecasting, uses multiple regression
techniques to "post-process" and analyze the results of an ensemble of different
models' forecasts.
Other meteorologists, among them Mr. Kummerow, at
Colorado State, agree that Mr. Krishnamurti has developed a potentially useful
technique. Mr. Krishnamurti creates his multimodel superensemble forecast by
using statistical techniques and the data on the models' past performances to
correct what he says are the collective forecast biases of the 11 global model
forecasts that make up a superensemble forecast.
"One model will have
better data and a better way of representing a lake and its physics, another
model could be better at representing a mountain, so the superensemble pulls the
best from all of them," Mr. Krishnamurti says. "The superensemble is the
collective wisdom of everybody, so to speak."
In a paper published in
Science last year, Mr. Krishnamurti and his colleagues describe how their
multimodel superensemble forecast made smaller errors in hindcasting the path
and intensity of dozens of tropical storms in 1997 and 1998 than did any one of
the model forecasts individually. "The superensemble outperforms all models,"
Mr. Krishnamurti says without hesitation.
Each computer model adds
unique and valuable information, Mr. Williford adds. The Florida State global
model, for instance, "has been tuned to the tropics," he says. The researchers
are now studying whether a superensemble can have an optimal number of
forecasts. They think, for instance, that each forecast has unique strengths and
that removing any one of them weakens the ensemble forecast. Using statistical
weighting techniques, the researchers try to minimize the effects of known flaws
in a particular model's forecasting skills.
Researchers from Mr.
Krishnamurti's group say they have also programmed their computer model to
simulate about 95 percent of the precipitation that can be detected by global
satellites, which further improved their forecasting accuracy. For a time, the
best they could do was 30 percent. When a global model simulates how certain
initial weather conditions will change during a period of, say, 24 hours, even
tiny errors in describing atmospheric conditions at the beginning of the model
run can grow out of control, "contaminating the forecast," Mr. Williford says.
Whatever the count of models, the more forecasts the researchers add to
the superensemble, the more computing power is required. It takes three hours
for the Florida global model, running on a small supercomputer with four I.B.M.
Power3 processors, to churn out a six-day forecast. Add more forecasts, and it
takes longer.
The researchers are still analyzing the results of their
experiments during this year's hurricane season, which was much less busy than
last year's, when Floyd disrupted many lives. And they are sending superensemble
forecasts twice a day to the National Hurricane Center in Miami for its
forecasters to evaluate.
Having always worked with limited computing
resources, weather researchers generally say they are used to accepting lower
thresholds of resolution, or realism, in their models, as a tradeoff for faster
computations. As Mr. Williford puts it, "Better computational capacity is only
going to help this process."
In the future, says Mr. Krishnamurti,
meteorologists working on faster computers may be able to apply the
superensemble concept to high-resolution, regional models for more objective and
accurate forecasts of the path and intensity of hurricanes.
"This is
coming, but it's maybe going to be another five years," he says.
http://chronicle.com
Section: Information Technology
Page: A51