On March 4, 1993 I visited Francis Balint, then Chief of the Automation Division for the National Meteorological Center (NMC) in Suitland, Maryland, USA. My notes on the meeting recently turned up. (n.b.: Many of the numbers quoted below have doubtless changed over the years.) Among the tidbits:
- A worldwide ballet: twice a day, at 0 hours and 12 hours Zulu (Greenwich mean time) people around the globe simultaneously collect data to share for numerical weather modeling --- pressure, temperature, wind speed and direction, etc.
- Disaster prevention is a big job for the meteorological service, from warnings of local tornadoes up to predictions of hurricane landfall locations.
- Numerical weather models generate a lot of output, maybe 1 GB/day; their inputs are in contrast quite small. Observations must be cleaned up and adjusted for consistency before they are put into the modeling process, to avoid artificial numerical instabilities --- things like unphysical shock waves that don't occur in real life. (see CookingTheBooks)
- The best US operational forecasts are made by a model called T126. It divides the atmosphere into 18 levels and 126 "waves" (Fourier components), equivalent to roughly 100 km grid cells. The European Centre for Medium-range Weather Forecasting, aka ECMWF, is using a T180 and is looking into even more detailed models.
- Forecast accuracy is measured by a "skill rate" parameter that quantifies the point where predictions break down into noise. The drop-off in the US happens 6-7 days out; it was 5 days before the 8-processor Cray Y-MP computer came into service. A request-for-proposals is out for an upgrade to the Y/MP-8.
- Good physics in a model counts for as much as raw resolution and computational power. It's hard on massively parallel machines to put in more physics ... things like variable heat loss due to radiation, effects of cloud cover, etc.
- "Ensemble modeling" is a key technique: tweak the input data and see how far into the future the output gets before it diverges. This tests stability, both of the model and of the atmosphere itself. Sometimes the weather is stable and one can predict farther ahead, as much as 10 days or so; other times, things are extraordinarily unstable and only short-term forecasting is possible. The key challenge: quantify the accuracy and limitations of one's predictions.
- The prime economic customers of weather forecasts:
- Agriculture --- a couple of days of advance notice of frost, rain, etc. can make a huge difference when bringing in crops
- Transportation --- especially aviation, but also boats, tankers, oil rigs, and to a lesser degree land transport
- Construction --- for scheduling outdoor projects, pouring concrete, etc.
(see ForecastFactory = http://www.his.com/~z/weather.html ...)
TopicScience - TopicPersonalHistory - Datetag20010908
(correlates: PureTheory, CookingTheBooks, StagesOfWork, ...)
![[Home]](/images/zh.gif)