Forecasting Trouble
A fire that destroyed a costly National Weather Service supercomputer in Maryland last month has undermined the accuracy of weather forecasts across the country while illustrating just how heavily modern forecasting depends on the ever-expanding brainpower of computers.
Especially during the last 10 years, the complex mathematical formulas that make up weather models, combined with more powerful computers, have let meteorologists spot the first breaths of Santa Ana winds or the early whirl of a fledgling hurricane days or even weeks in advance.
But many of the nation’s most advanced weather models collapsed into a mere heap of equations Sept. 27 when an electrical fire broke out inside a Cray C90 supercomputer in Suitland, Md. The computer processed a dozen National Weather Service models, including those that foresee hurricanes and other hazardous weather.
Firefighters sprayed the burning computer with dry chemicals that corroded its insides and destroyed it, said Wayman Baker of the Weather Service’s National Centers for Environmental Prediction.
“Because it was providing the central forecast guidance for the Weather Service, it was playing a critical role at the time of the fire and it’s hard to put a value on that,” Baker said. “What we lost was our ability to execute the models on that computer, which is critical to what we do.”
When the Weather Service purchased the computer for $45 million in 1994, it was the most powerful system on the market, capable of 6 billion calculations per second. It gobbled weather observations from around the world and tied them together with equations.
From these it fashioned images revealing what a weather system would do one day later, a day after that, and many more days after that--up to 16 days in the future. “It’s a powerful tool that transforms data into models of the weather that cover the entire globe,” said Carlos R. Mechoso, a professor of atmospheric sciences at UCLA.
“In terms of forecasting, computers have really brought us to where we are today, and one of the biggest challenges is keeping them up to date,” Mechoso said.
Though the Weather Service hopes to have a replacement computer up and running soon, it is currently struggling with inferior equipment.
After the fire, the prediction center turned to two backup computers with about 40% of the capacity of the Cray. Because they can model weather only 10 days ahead, the Weather Service’s Climate Prediction Center recently notified its field offices that “there is no objective basis for making forecasts at the 8- to 14-day range, and no further messages of forecasts will be issued until model guidance at that range again becomes available.”
The alternate computers also cannot extract as much detail from data collected by airplanes flying through hurricanes, said Jerry Jarrell, director of the National Hurricane Center in Miami. “Is that degrading our forecasts?” he asked. “Nobody in the world knows, including us.”
Navy computers took over some hurricane modeling after the fire, but weather models continue to arrive late, forcing forecasters to issue predictions based on scantier information, Jarrell said. “That has had a fairly big impact, but the result is the same: You’re still making an objective forecast based on the best available data,” he said. “It’s like a doctor that no longer has access to some tests. You don’t know if the tests would have helped, because you don’t know what they would have told you.”
The Cray churned through one primary weather model for North America four times a day, giving forecasters updated pictures of future weather every six hours. The slower backup duo processes the same model only twice a day at less than half the resolution.
*
Using the fuzzier model of forthcoming weather is like watching a football game through low-power binoculars: It’s tougher to spot patterns that may alter the eventual outcome. “We’re not seeing some of the smaller details we had begun to rely on as clues,” said Paul Knight, a meteorology instructor at Pennsylvania State University.
At AccuWeather Inc., which translates National Weather Service models into forecasts for newspapers nationwide, meteorologist Mark Tobin said the models may blur compact weather developments such as thunderstorms and snow flurries, “so the level of confidence in the forecasts is not as high.”
Cold fronts sweeping into Wyoming in recent weeks have arrived six to 12 hours before the models showed they would, “wreaking havoc on some of our predictions,” said Jeffrey Williams of DayWeather Inc., a private weather service in Cheyenne, Wyo.
Today’s escalation in computing power, however, has eased the Weather Service’s crisis somewhat. Navy and Air Force computers have begun handling some models. U.S. forecasters have also drawn more heavily on independent European and Canadian weather models, all now publicly available on the Weather Service’s Los Angeles Web site, https://www.nwsla.noaa.gov/forecast.html.
“Models have multiplied so dramatically over the last few years that this hasn’t been as damaging as it would have been, say, 10 years ago,” said Gary Ryan, a forecaster in the National Weather Service office in Oxnard. “Knowing the strengths of the various models is where the forecaster’s skill comes in.”
The Weather Service had already contracted for a $36-million IBM supercomputer to replace the Cray and now hopes to rush it into service next month, Baker said. The IBM will eventually boast 29 times the processing power of the Cray, which was near the end of its useful life as a weather watcher even before it burned up.
