|(Click on the above images to view the full displays.) The Sydney area was pelted by marble-size hail and 2 inches (5 centimeters) of rain on 22 September 1999, during tests of the RAP Auto- nowcaster. At left is the Doppler radar output for 0930 GMT (about 7:30 p.m.. Sydney time) showing reflectivities above 30 dBZ, a common threshold for thunderstorms. In the image at left, the western (leftmost) storm, directly over the main Olympic site, developed and arrived after a larger storm complex had moved to the east. At right is the Auto-nowcaster's forecast issued at 0859 GMT for 0929 GMT. Although the western storm had not yet formed at 0859, the Auto-nowcaster correctly predicted its development. (Illustrations courtesy RAP.)|
While athletes from around the world compete for the gold in this month's Olympic Games in Sydney, an elite team of automated forecast tools is going head to head with Australia's weather. Human forecasters will use software furnished by Australia, Canada, Great Britain, and the United States to help provide short-term Olympic forecasts from 15 September to 1 October. It's part of a three-month exercise to see how much of a boost the automated tools can provide to flesh-and-blood forecasters.
RAP's Auto-nowcaster was installed in Sydney last fall for testing. The five-year-old system uses 30 different algorithms to predict the initiation, growth, and decay of thunderstorms. Outlooks are issued every five minutes for periods of up to an hour. Joining the Auto-nowcaster this fall are four other automated systems. "Each one does a little something different," says Jim Wilson, RAP/ATD senior scientist and project manager for the Auto-nowcaster.
Forecasters from Australia's Bureau of Meteorology will issue official Olympic forecasts by mixing and matching output from the five automated systems and applying their own insight. Their outlooks will go to emergency managers, venue managers at the Olympics, and flight controllers at Sydney Airport. The official forecast site can be accessed on the Web.
Joining the Auto-nowcaster are:
The experiment will draw on data from two Doppler radars, three wind profilers, about 20 weather stations, and four radiosonde launches a day from Sydney Airport. "Unfortunately, satellite data are available only once an hour," says Jim. That's because a satellite-bearing rocket from Japan exploded on takeoff several months ago. The once-hourly data "isn't frequent enough for us to do our best, especially when we're trying to predict new storms. The satellite adds a lot to being able to see the first cumulus clouds."
Most of the Auto-nowcaster's initial development was funded by the FAA and the U.S. Army, along with some core funds from NSF. The software has proven its mettle in tests at the NWS office in Sterling, Virginia, and the Army's White Sands Missile Range (New Mexico) and Redstone Arsenal (Alabama). Recent improvements have come through support from the U.S. Weather Research Program. Eventually the NWS plans to bring Auto-nowcaster concepts into its radar analysis routines, but "it all has to wait for open computer [platforms]," says Jim. "They need more computing power."
By tracking convergence lines (gust fronts, sea breezes, and other zones where air masses collide), the Auto-nowcaster anticipates where the next storm might form. "The idea is to do better than if you just extrapolate the existing storms," explains Jim. Software engineers Jaimi Yee, Niles Oien, and Nancy Rehak and associate scientist Dan Megenhardt have enhanced the Auto-nowcaster in recent years. Jenny Sun and Andrew Crook (MMM/RAP) furnished advanced modeling capabilities that allow wind fields to be deduced from a single Doppler radar.
"We've made some major strides in the last couple of months in our ability to forecast growth and decay," Jim says. For instance, the system can now analyze where and when convergence lines will run into existing storms and help stimulate their growth. Someone could watch radar closely and spot this process, but often there's too much going on at once for a forecaster to monitor every possible interaction. Back in the 1980s and early 1990s, when RAP was first developing and testing techniques for very short period forecasts, "it didn't take us long to realize that the human wears out," says Jimthus, the creation of the Auto-nowcaster.
Computerizing the process isn't easy, though: "To get a machine to see what your eye can see is extremely difficult." RAP and the Massachusetts Institute of Technology's Lincoln Laboratory have been working for many years to develop an algorithm to automatically detect convergence lines observed on radar. Led by software engineer Dave Albo, they've settled on a technique to be tested in Sydney in which a person will input the convergence line's location and the Auto-nowcaster will do the rest. As Jim puts it, "We've taken the approach of having a human help the machine."