Beating the Pros at the Weather Game:

FSU "Superensemble" Approach to Predicting Hurricanes' Path

    Have weather scientists at FSU developed a way to predict the path of hurricanes better than what you hear on the evening news? A look at the track record of the '99 hurricane season suggests just that.
    Beginning in August 1999, meteorologist Dr. T.N. Krishnamurti and his staff trained their computers onto 15 named storms in the Atlantic Ocean. Near the season's end in late November, the FSU team posted better numbers, in terms of percentages of more accurate calls on actual storm tracks, on most of them.
    The best call? Krishnamurti's lab outperformed all other models for the track of Hurricane Floyd, the season's most destructive storm, accurately predicting its swing around the Georgia and South Carolina coast and, four days out, predicting its Sept.16 landfall at Cape Fear, North Carolina. By comparison, the official forecast from the National Hurricane Center predicted Floyd would sweep perilously close to Florida's northernmost beaches before slamming ashore just south of Charleston, South Carolina.
    As a result, governors of Florida, Georgia and South Carolina issued orders to coastal communities that triggered the largest peacetime evacuation in history, according to federal officials. In Florida alone, 1.3 million people fought to get out of harm's way, jamming all major highways out of Jacksonville and surrounding areas. The Georgia and South Carolina exodus saw thousands of motorists stuck for hours in stalled interstate traffic that stretched for miles inland.
    In the storm's aftermath, CBS news anchor Dan Rather ran a story that called into question the necessity of this massive evacuation based on what forecasters really knew about Floyd's behavior. For background information on the piece, Rather's news team used, without attribution, some of the results of Krishnamurti's calculations. The FSU findings can be viewed as supporting arguments that state and federal officials may have overreacted to the emergency. Rather emphasized that disaster preparedness officials constantly worry that bad or hasty predictions reinforce public reluctance to obey evacuation orders, thereby compromising compliance with such orders in the future and potentially opening the door to catastrophe.
    News of his findings kept Krishnamurti and his staff almost as busy handling media requests last fall as did their ongoing research. After a write-up of the research appeared in a September issue of Science, CBS reporter John Roberts featured the FSU lab on the CBS Evening News. Roberts led with the story of lab's successful prediction for Hurricane Dennis--that the storm wouldn't make landfall, an on-the-money call that came nearly six days before federal Dennis-watchers concluded the same thing. ABC's Peter Jennings sent a camera crew to campus in October, and dozens of lesser lights showed interest as well in what could be a revolutionary breakthrough in the terribly tricky game of predicting the path of hurricanes.
    Krishnamurti's success as a big storm prognosticator comes in the third year of a research project never before attempted in hurricane science. He credits his seven-member team of graduate students and post-docs with developing the key to the project's success--a brilliant new method of computer-crunching massive amounts of weather data.
    As Krishnamurti assistant Eric Williford explains it, traditional hurricane forecasting is based on numerical models created for computers to digest enormous streams of data collected from dozens of land, sea, sky and space-based sources. But every hurricane model ever built comes with built-in errors, or biases, says Williford, which throw the final results off in varying degrees.
    FSU's novel approach is to combine--with the aid of three in-lab supercomputers--every model the team can get its hands on and run them through a clever "de-biasing" set of algorithms that eliminate or suppress known quirks in the data. The result is an ensemble of models, or in the parlance, a "superensemble" that coalesces the best aspects of the sum of its parts and spits out as close to a bias-free forecast as exists.
    Still in the testing stage, the superensemble model has nonetheless shown remarkable promise in its first full season of hurricane plotting, says Krishnamurti. He's seen a 20- to 30-percent improvement overall in his own forecasting and up to 100 percent in other models using the technique. Support from a private company--an insurance underwriting consortium--already has contributed towards the main computers, an IBM SP supercomputer and two IBM high-powered workstations, and if tests continue to show promise, interest in getting the technique into the public arsenal of forecasting tools surely will follow.
    The latter development could prove to be a boon to regions outside hurricane alley, meaning most of the continental U.S. Given enough computing power, Krishnamurti says the model could be applied to everyday weather analysis, a prospect that could make today's much-ballyhooed "six-day forecasts" of nightly weathercasts finally worth paying attention to.