In a paper presented to the 18th IIPS Conference (2002), the present author described some early work towards the development of a knowledge-based system for the generation of internet weather forecasts. The performance of the 2002 system was evaluated (real-time) over one month (November 2001) and found to be superior to forecasts based on persistence and climatology. However, the system's predictions were somewhat inferior to the corresponding officially issued forecasts, especially for days 1 and 2. For more details, readers may refer to: http://www.weather-climate.com/internetforecasts.ppt

A subsequent analysis of the verification data suggested that there were three main reasons for this inadequate performance:

• Firstly, the forecasts were largely a function of the broad-scale synoptic pattern suggested by the NWP models;
• Secondly, the system failed to take into account NWP model forecasts of the thermal pattern; and,
• Thirdly, the system failed to take into account NWP model forecasts of moisture in the troposphere.

The current paper describes how the knowledge-based system has been modified in order to address the aforementioned deficiencies. It also presents the results of a verification exercise on developmental data. This exercise demonstrates that potential for a significant increase in the accuracy of the forecasts has been achieved.

The 2002 system operated by producing its predictions from a restricted number of discrete "forecast sets". The set that was chosen (by the system) was largely determined by the particular synoptic pattern suggested by the selected NWP model. The 2003 modification utilises regression analysis to allow predictions to be selected from a continuous array of possible forecasts. The particular form of regression analysis employed is parameter enveloping. Parameter enveloping allows definition of how the various predictors impact upon, or envelope, the influence (on a predictand) of other predictors.

The 2003 modification utilises 40 years of data (1961-2000), made up of 14,610 individual synoptic situations, in its development. The data is stratified into a set of 50 synoptic types, utilising NCEP data and a synoptic-typer interface (refer to another paper in the current session by Robert Dahni). Regression analysis is then carried out on data associated with each of the synoptic types. Maximum temperature forecasts for Melbourne (generated from the developmental data set) display greater accuracy than that achieved historically by official day-1 predictions, for every one of the synoptic types.

However, it should be pointed out that, over the years, there has been an increase in the accuracy of the official day-1 temperature forecasts.

It is planned to carry out an evaluation of the performance of the modified system real-time, and over one month, in November 2002, and to compare its performance with that of the "pilot" system in November 2001. Readers interested in detailed results of this evaluation may refer to: http://www.weather-climate.com/internetforecasts.html