A "pilot" knowledge-based system for the automated generation of internet weather forecasts is described. The system has been developed for the small (227,000 sq km) southeast Australian State of Victoria. The categories of weather forecasts generated include forecasts for public, aviation and marine interests.

At the core of the system is an algorithm, written in JavaScript. The algorithm combines a statistical interpretation of the NWP model data in terms of local weather, with other knowledge. The output generated is HTML code, which is then up-loaded to a WEB Site.

The statistical interpretation component of the system identifies the type of synoptic pattern suggested by the NWP model output, suites of forecasts having been derived for each of the different synoptic patterns. Other knowledge about weather associated with the different synoptic patterns has been utilised in the development of the suites of forecasts.

One of the major benefits of the system is the elimination of typing on the part of forecasters. This would, thereby, allow for the production of forecasts for a greatly increased number of localities than currently. The system provides precis forecasts (worded, and also depicted by an icon graphic), predictions of maximum and minimum temperature, and precipitation (probability and amount), all out to day seven for 210 places. At present, worded forecasts, and forecasts of maximum and minimum temperature, are provided officially only out to day four, and only for 24 places.

The performance of the system is evaluated utilising five forecast accuracy measures. These are: (1) the root mean square error of the minimum temperature predictions; (2) the root mean square error of the maximum temperature predictions; (3) the root mean square error of the predictions of precipitation amount; (4) the percentage correct precipitation/no precipitation forecasts; and, (5) the Brier score of the probability of precipitaion forecasts.

Preliminary verification statistics are encouraging. They show that, on each of the five measures, the system's performance is superior to forecasts based purely on climatology or persistence. However, the system is presently inferior to the independently produced official forecasts.

The system's skill would be expected to increase as new knowledge is incorporated into its operation. This is very much along the lines of Ramage's proposed "iterative" approach to "locking in" improvements in forecasting methodology, as described in his 1993 BAMS paper, and subsequently illustrated by the present author, in his 1996 Ph.D. thesis.

The system's approach has some similarities to that of the US Interactive Forecast Preparation System (IFPS). However, there are two key differences -

(1) One difference is that in operation, the system's forecasts are intended to be (mostly) automatically generated and transmitted. The main interaction that takes place is in the context of utilising forecast verification analyses (after the event) to iteratively incorporate additional forecaster knowledge into its algorithm.

(2) The other difference lies in its provison of forecasts for specific localities, rather than for an array of grid points for subsequent interpretation by private providers (as does IFPS). Provision of forecasts for specific localities might be the preferred route for a country such as Australia, where there are only a small number of private meteorologists.