Producing a weather forecast

How computers predict weather (continued)

The present system enables temperatures, winds, moisture and pressure to be analysed simultaneously over a wide area. Increased computing power has brought an increase in the predictions from 36 hours ahead to twice-daily predictions eight days ahead. These far more sophisticated models have significantly improved the timeliness of forecasts and warnings. Today's forecasts four days ahead are as accurate as 24-hour forecasts were a decade ago.

In addition, 'limited area' models with a grid size of 12km are now 'nested' within the large-area models to provide short-term forecasts ('nowcasts') over south-west and south-east Australia.

These are valuable for predicting severe local storms and flooding, including forecasting detailed weather for important events such as the Sydney Olympics. They help predict dispersion of bushfire smoke, and city air quality.

To cover the vast oceans around Australia, where few observations are available, the computers are also fed 'pseudo observations' derived from manual analyses and satellite interpretation.

Weather modellers must also confront the basic problem of representing a meteorological field over a rectangular array of grid points, given that the observations are not at the grid points. Observations made close to grid points must be allocated a greater influence than more distant observations.

The modellers start by deriving an initial 'first guess' field by giving significant weights to a large number of observations. This field is refined by adjusting the scale of influence of observations, assigning negligible value to observations distant from the grid point.

Because many meteorological parameters are related, we can do better than simply using one parameter at the different points in space. For instance, when analysing the surface wind it would make sense to incorporate the surface pressure data as well.

Data are smoothed by a filter process which averages data over several grid points. The most efficient filtering demands the largest possible number of grid points and very high supercomputer capacity.