Home Forecasts, Warnings, Flood Warnings, Charts, Observations, Marine Services, Satellite Images Maps, Rainfall Outlook, Climate Averages, Climate and Our Environment Flood Warning Service, Water Resources Assessment, Hydrometeorological Advisory Service Our Products, Services and Access Methods Registered User information and LOGIN Phone Number, E-mail, Addresses and Postal Addresses Bureau Contacts, Charter, Organisation, Programs, WMO Media Releases, Conferences, Events Publications Educational Brochures, Library, School Projects, Career Information Feedback Search the site for words or phrases Help - FAQs, Glossary

Navigation Bar Graphic

PROGRESS ON A KNOWLEDGE-BASED INTERNET WEATHER FORECASTING SYSTEM by Harvey Stern [updated paper for delivery to the 19th IIPS Conference]

(Link to 19th IIPS Conference Program)
(Link to 19th IIPS Conference Presentation)
(Link to 19th IIPS Conference System)
(Link to updated 18th IIPS Conference Paper)
(Link to 18th IIPS Conference Presentation)
(Link to abbreviated version of the 18th IIPS Conference System)
(Link to Background Information about the Author)

ABSTRACT

An early version of the knowledge-based system was presented to the 18th IIPS Conference (Stern, 2002a&b). The system was developed for the small (227,000 sq km) southeast Australian State of Victoria. It was described as being capable of generating forecasts for public, aviation, marine and media interests, in languages other than English, and for more than 200 localities in Victoria.

A major benefit of a knowledge-based system is that it incorporates an extensive "bank" of forecaster experience. Ramage (1993) has proposed an "iterative" approach to "locking in" improvements in forecasting methodology. The system's skill increases as new knowledge is incorporated into its operation. Hence, progress is gradually made towards the realisation of Ramage's dream. The system is (therefore) not seen as "yet another" instrument of forecast guidance. Rather, its development is seen as a logical step along the path of having the computer replicate the best aspects of the manual side of the forecast process, by systematically "locking in" new knowledge. As Brooks (1995), wrote: "technology, which initially allowed humans to make routine weather forecasts, will soon close that avenue of human endeavour ... (and thereby permit) concentration on severe events".

The system's forecast performance (during November 2001) was evaluated for the city of Melbourne. The evaluation showed that, although superiority over climatology was achieved, the forecasts (on most measures) proved to be inferior to the official forecasts.

The system was then "scaled back", new knowledge added, and re-evaluated over a 100-day trial (Stern, 2003). The hope had been that its deficiency (in comparison with the official forecasts) could be somewhat reduced. As was the case with Noone and Stern's (1995) verification of the first 100 days of rainfall forecasts from the high resolution version of the Australian Bureau of Meteorology's Global Assimilation and Prognosis System (GASP), an evaluation period of 100 days duration is considered to be of sufficient length to yield a useful assessment.

For the first 4 days, using a skill score that combines all features of a forecast, that deficiency appeared to have been not only reduced, but, in fact, eliminated. However, a skill score that combines all features of a forecast disguises individual components. Closer examination of the verification data reveals that, although substantial progress towards achieving computer replication of the manual forecast process has been made, this is uneven. For example, while the performance of the system's temperature forecasts is either equal to, or better than, corresponding official predictions for the first 5 days, in regard to the precipitation forecasts, this is only the case for Day-1.

The performance of the system over the 100 days is so impressive, that one must entertain the possibility that it was simply a "fluke". For this reason, further testing needs to be undertaken. Also, an investigation of the likely cause of the deficient performance at forecasting precipitation beyond Day-1 needs to be carried out, with the consequent incorporation of any new knowledge that such an investigation suggests. Should the further testing confirm the 100-day trial results, we could see ourselves at the "dawn" of the operational implementation of Ramage's approach to weather forecasting which, through computer replication of the manual process, would allow for:

  • Systematic incorporation of new procedures that may lead to a quantum leap in the accuracy of the forecast products; and,
  • The opportunity to greatly increase the number and variety of such products.

The author briefly diverts from the main aim of this work, examining the verification data associated with experimental official (seven day) forecasts, with a view to exploring whether it might now be time to issue these seven day forecasts to the public. Such examination suggests that, indeed, it might now be time.

Table of Contents

ABSTRACT


  1. Introduction
  2. Background
  3. Describing the System
    1. Synoptic Typing
    2. Standardised Terminology
    3. "Bank" of Forecaster Experience
  4. 2001 evaluation
    1. Analysis of November 2001 trial
    2. Modifications
  5. Specifications for 2002/2003 trial
    1. Deducing precipitable water
    2. Climate Change Implications
    3. Verification on developmental data
    4. Purpose of 2002/2003 work
  6. Example
  7. Summary of the 2002/2003 evaluation
    1. Performance over Day-1 to Day-4
    2. Replicating the forecast process
    3. Performance for Day-1
    4. Performance using the perfect prog assumption
    5. Note on the 2002/2003 evaluation
  8. Public forecasts out to 7 days?
  9. A Forecasting Research Tool
    1. Illustration 1
    2. Illustration 2
  10. Potential
  11. Concluding remarks
  12. References

APPENDIX

  1. Postscript
  2. An Alternative Approach

APPLYING THE SYSTEM
COPYRIGHT NOTICE

Table of Figures

  • Map 1 Map depicting the location of the southeast Australian State of Victoria.
  • Map 2 Map depicting the location of places referred to in the text.

  • Fig. 1 Segment of the output of the 18th IIPS Conference version of the system,
    illustrating the aviation component.
  • Fig. 2 Segment of the output of the 18th IIPS Conference version of the system,
    illustrating the media graphics capability.
  • Fig. 3 The web-based product that generates time series of expected weather.
  • Fig. 4 Illustration of the overall accuracy of the forecasts in 2001.
  • Fig. 5 An illustration of parameter enveloping.
  • Fig. 6 Precipitable water/850 hPa temperature relationship.
  • Fig. 7 Long-term upward trend in Melbourne's minimum temperature
    (30-year period ended in indicated year).
  • Fig. 8 Long-term upward trend in Melbourne's maximum temperature
    (30-year period ended in indicated year).
  • Fig. 9 Accuracy of developmental data forecasts.
  • Fig. 10 NWP MSL prognosis.
  • Fig. 11 NWP 700 hPa relative humidity prognosis.
  • Fig. 12 NWP 850 hPa temperature prognosis.
  • Fig. 13 The system's output, after inserting data from the NWP prognosis into the form.
  • Fig. 14 Corresponding official forecast.
  • Fig. 15 RMS Error of the QPFs.
  • Fig. 16 Percentage Correct Rain/No Rain Forecasts.
  • Fig. 17 Brier Score of the PoP Forecasts.
  • Fig. 18 Reliability Diagram for the Day-1 to Day-6 PoP Forecasts.
  • Fig. 19 Reliability Diagram for the Day-1 PoP Forecasts.
  • Fig. 20 ROC Diagram for the Day-1 to Day-6 PoP Forecasts.
  • Fig. 21 ROC Diagram for the Day-1 PoP Forecasts.
  • Fig. 22 Illustration of the accuracy of the minimum temperature forecasts.
  • Fig. 23 Illustration of the accuracy of the maximum temperature forecasts.
  • Fig. 24 Illustration of the overall accuracy of the temperature forecasts.
  • Fig. 25 Illustration of the overall accuracy of the forecasts.
  • Fig. 26 Anomaly Correlation for ECMWF and GASP models
    (compare with Fig. 25).
  • Fig. 27 The extent that official maximum temperature forecasts (squares) for
    seventeen Victorian centres, issued several days ahead,
    improved upon the guidance (diamonds) - from Stern (1998).
  • Fig. 28 The extent that official maximum temperature forecasts (squares) for
    the city of Melbourne alone, issued several days ahead,
    improved upon the guidance (diamonds) - from Stern (1998).
  • Fig. 29 Synoptic Type 41 (refer to Dahni, 2003).
  • Fig. 30 The Annual March of the PoPs.
  • Fig. 31 The Annual March of the QPFs.
  • Fig. 32 The Annual March of the PoPs should one add cyclonicity to the predictors.
  • Fig. 33 Synoptic Type 32 (refer to Dahni, 2003).
  • Fig. 34 The Annual March of the Maximum Temperature Forecasts.
  • Fig. 35 The Annual March of the Maximum Temperature Forecasts should one add
    north (MSLP Gabo-Gambier) to the predictors.
  • Fig. 36 Segment of the output of the 18th IIPS Conference version of the system,
    illustrating the Chinese component.
  • Fig. 37 Segment of the output of the 19th IIPS Conference version of the system,
    illustrating the Chinese component.
  • Fig. 38 Synoptic Type 14 (refer to Dahni, 2003).
  • Fig. 39 The Annual March of the Maximum Temperature Forecasts for Synoptic Type 14.
  • Fig. 40 Illustration of the system's alerting capability.
1. Introduction

An early version of the knowledge-based system was presented to the 18th IIPS Conference (Stern, 2002a&b). The system was developed for the small (227,000 sq km) southeast Australian State of Victoria (refer now to Map 1 and Map 2) and was described as being capable of generating forecasts for public, aviation (Fig. 1), marine and media interests (Fig. 2), in languages other than English, and for more than 200 localities in the State.

2. Background

Most work that has been carried out in Australia towards providing automated forecast guidance has focused upon weather elements contained in public forecasts (Stern, 1980, 1996, 1997, & 1999b; Stern et al., 1984a&b, 1987; Dahni, 1988; Dahni and Stern, 1995; Dahni et al., 1984a&b). Little such work has been carried out in Australia towards providing guidance for aviation forecasting, although Stern and Parkyn (1999 & 2001) have developed a web-based forecasting technique for Melbourne Airport, and work progresses towards automated generation of terminal aerodrome forecasts (Godau et al., 2001). The Australian Bureau of Meteorology has a web-based product that generates time series of expected weather (Fig. 3), but this product is comes directly (without modification) from the latest NWP model output.

Brunet et al. (1988) found that Perfect Prog (PP) forecasts perform better than MOS forecasts for short-term predictions and note that PP forecasts possess the overwhelming advantage of portability of the system when the driving model changes. For these reasons the PP approach is used in the development of the system described herein.

The large volume of output proposed here (forecasts for than 200 localities) would be difficult (if not impossible) to generate manually, but would be straightforward (to generate) utilising automated systems.

3. Describing the System

At the core of the system is an algorithm, written in JavaScript. The algorithm combines a statistical interpretation of NWP model data in terms of local weather, with other knowledge. The statistical interpretation component of the system involves identifying the synoptic pattern type suggested by the NWP output, suites of prediction equations and forecasts having been derived for each of the different synoptic patterns. Knowledge about the weather associated with each synoptic type is utilised in developing these suites of equations and forecasts.

3.1 Synoptic Typing

The synoptic typing procedure used by the system is that for southeast Australia first referred to by Treloar and Stern (1993), and defined in detail by Stern and Parkyn (1999). The synoptic types are described in terms of the direction, strength and curvature of the isobaric (surface) flow. These characteristics are determined from a grid of pressure values as follows:

3.2 Standardised Terminology

Nearly three decades ago, Dobryshman (1972) discussed the importance of a standard forecasting terminology. A standard terminology, relatively straightforward to implement in the context of an automated system, ensures unambiguous interpretation of the forecasts, and also allows the forecasts to be objectively verified. To this end, the system employs:

3.3 "Bank" of Forecaster Experience

A major benefit of the knowledge-based system is that it incorporates an extensive "bank" of forecaster experience. Ramage (1993) has proposed an "iterative" approach to "locking in" improvements in forecasting methodology, and this approach was recently illustrated by the present author (Stern, 1996 & 1999b). Indeed, the automated nature of the system lends itself to Ramage's approach, and this is the approach adopted by the forecasting system described herein. The system's skill increases as new knowledge is incorporated into its operation. Hence, progress is gradually made towards the realisation of Ramage's dream. The system is (therefore) not seen as "yet another" instrument of forecast guidance. Rather, its development is seen as a logical step along the path of having the computer replicate the best aspects of the manual side of the forecast process, by systematically "locking in" new knowledge. As Brooks (1995), wrote: "technology, which initially allowed humans to make routine weather forecasts, will soon close that avenue of human endeavour ... (and thereby permit) concentration on severe events".

The system's main human-interaction is in utilising forecast verification analyses (after the event) to iteratively incorporate new additional forecaster knowledge into its algorithm.

4. 2001 evaluation

The system's forecast performance (during November 2001) was evaluated for the city of Melbourne. The evaluation showed that, although superiority over climatology was achieved, the forecasts (on most measures) proved to be inferior to the official forecasts. Readers may now wish to refer to Fig 4.

The system was then "scaled back" (for example, the rural city of Mildura, in northwestern Victoria, the alpine town of Mt Hotham, in northeastern Victoria, and the outer suburb of Watsonia, were now the only localities with forecast information aside from Melbourne).

The results of the November 2001 trial were analysed, potential improvements in the forecasting process (employed by the system) were identified, and those improvements were "locked in". With the new knowledge added, the system (so modified) was then subjected to another evaluation for the city of Melbourne (Stern, 2003a&b), this time over 100 days (between 25 September, 2002, and 2 January, 2003, for Day-1 forecasts). As was the case with Noone and Stern's (1995) verification of the first 100 days of rainfall forecasts from the high resolution version of the Australian Bureau of Meteorology's Global Assimilation and Prognosis System (GASP), an evaluation period of 100 days duration is considered to be of sufficient length to yield a useful assessment.

4.1 Analysis of November 2001 trial

An analysis of the November 2001 verification data suggests three main reasons for inadequate performance:

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.

4.2 Modifications

The 2001 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 2002/2003 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 (Stern, 1994a&b; 1996 & 1999b). Parameter enveloping allows definition of how the various predictors impact upon, or envelope, the influence (on a predictand) of other predictors.

The process of parameter enveloping provides a means to explicitly interpolate between elements in a data set. In its simplest form, it may be illustrated as follows:

An illustration of this process is presented in Fig. 5.

5. Specifications for 2002/2003 trial

Terms in the maximum temperature regression equations include combinations of:

Terms in the minimum temperature regression equations include combinations of:

Terms in the probability of precipitation equations, and also the amount of precipitation equations, include combinations of:

In operation, if the solution to the probability of precipitation equation is <50%, then, regardless of the solution to the amount of precipitation equation, that amount is ''set" equal to "nil".

5.1 Deducing precipitable water

Our global prediction system provides forecasts of relative humidity at 700 hPa (RH), and temperature at 850 hPa (T), but not (directly) of total precipitable water (W). Fig. 6 depicts the relationship between precipitable water and the temperature at 850 hPa at Melbourne. It demonstrates that the relationship between the precipitable water of a fully saturated atmosphere and the 850 hPa temperature is approximately:

W=20+1.5T (3)

Hence, cognisant of the facts that:

one might assert that the relationship between W and T and RH would be approximately:

W=(max(MIN(RH,90%),30%))x(20+1.5T) (4)

Equation (4) may, therefore, be used to deduce the NWP model forecast of total precipitable water.

5.2 Climate Change Implications

The modification takes into account the observed long-term upward trend in temperature (Fig. 7 and Fig. 8) which is due to a combination of overall global warming and local urbanisation.

Firstly, 0.2 deg C is added to the forecast minimum temperature, because the 1991-2000 mean is 0.2 deg C above the 1961-2000 mean - the assumption being that the 1991-2000 data better represents the current situation. Secondly, 0.3 deg C is added to the forecast maximum temperature, because the 1991-2000 mean is 0.3 deg C above the 1961-2000 mean.

Subsequent evaluation of the system's forecast temperatures confirms the need to take into account the upward trend in temperature. The mean bias (forecast-observed) of the Day-1 forecasts of minimum temperature was -0.09 deg C, while the mean bias of the Day-1 forecasts of maximum temperature was -0.21 deg C.

That a negative bias (albeit small) was observed, in spite of the adjustment, may be a reflection of the ongoing nature of the upward trend.

5.3 Verification on developmental data

The 2002/2003 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 (as defined by Treloar and Stern, 1993; Stern and Parkyn, 1999), utilising NCEP data and a synoptic-typer interface (Dahni, 2003). 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 (see Fig. 9). This may be deduced because all of the diamonds lie below the "no-improvement" line of circles, thereby suggesting that for all synoptic types, the developmental forecasts are better than the historical forecasts. 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. To illustrate, the RMS error of day-1 maximum temperature forecasts over the last 10 years, 1991-2000, was 2.23°C, in comparison with 2.69°C for the entire 1961-2000 period (Dawkins and Stern, 2003).

5.4 Purpose of the 2002/2003 work

The purpose of the current work is twofold:

6. Example

The charts following present an example of the system's forecast. Fig. 10 depicts the MSL prognosis, Fig. 11 depicts the 700 hPa relative humidity prognosis, and Fig. 12 depicts the 850 hPa temperature prognosis. Fig. 13 depicts the system's output, after inserting data from the NWP prognosis into the form. Fig. 14 depicts the corresponding official forecast.

7. Summary of the 2002/2003 evaluation (refer to note at Section 7.5)

Day-1 Internet Forecasts for rainfall (as measured by the RMS error of the QPFs (Fig. 15), % correct indications of "rain" or "no rain" (Fig. 16), and the Brier Score (Brier, 1950) (Fig. 17), are slightly better than the official forecasts. Day-2 Internet Forecasts for rainfall are inferior to the official forecasts on most measures (the exception is the "rain/no rain" measure), but are better than climatology, while Day-3 to Day-6 Internet Forecasts for rainfall display a level of skill that is little different to climatology on most measures (the exception is the "rain/no rain" measure).

There does not appear to be a significant bias in the Internet Forecasts. Over the 6x100 days of the trial:

The reliability diagram for the Probability of Precipitation (PoP) forecasts, prepared utilising data from Day-1 to Day-6 combined (Fig. 18), is typical of an "over-confident" set of forecasts. This is not unexpected, given that the system of Internet Forecasts is based on the assumption that the prognoses are "perfect".

The reliability diagram for the Probability of Precipitation (PoP) forecasts, prepared utilising data from Day-1 only (Fig. 19), shows a somewhat improved calibration (except for the data point representing PoPs of between 60% and 80%).

The Relative Operating Characteristics (ROC) diagram for the Probability of Precipitation (PoP) forecasts, prepared utilising data from Day-1 to Day-6 combined (Fig. 20), suggests the forecasts do display some skill. This is because the area under the curve, which is 0.687, is greater than 0.5 (the "no-skill" case). However, the skill is not regarded as "meaningful" according to a definition by Stanski et al. (1989), who require the area to be greater than 0.7 to be considered "meaningful".

However, the ROC diagram for the Probability of Precipitation (PoP) forecasts, prepared utilising data from Day-1 only (Fig. 21), suggests that these forecasts do display "meaningful" skill, the area under the curve being 0.753.

Day-1 Internet Forecasts for minimum temperature (as measured by the RMS error of the forecasts - Fig. 22) are slightly worse than the official forecasts, Day-2 and Day-3 Internet Forecasts for minimum temperature are slightly better, Day-4 Internet Forecasts are slightly worse, while Day-5 Internet Forecasts are slightly better. Day-6 Internet Forecasts for minimum temperature are worse than climatology.

Day-1 to Day-4 Internet Forecasts for maximum temperature (as measured by the RMS error of the forecasts - Fig. 23) are slightly better than the official forecasts. Day-5 Internet Forecasts for maximum temperature are slightly worse than the official forecasts, while Day-6 Internet Forecasts for maximum temperature are worse than climatology. That there is little improvement, in the official maximum temperature forecasts, between Day-6 and Day-3, may be a consequence of forecaster reluctance to change a prediction (so far in advance) once it has been decided upon.

Overall, between days 1 and 5, the accuracy of the Internet Forecasts for temperature are as good as, or even slightly better than, the official forecasts for temperature (Fig. 24).

7.1 Performance over Day-1 to Day-4

The Combined Skill Score (CSS), which has previously been referred to by Stern (1998, 1999, 2002), combines verification measures such as the Brier Score (on PoP estimates), performance at predicting rain or no rain, the RMS error of the QPFs, and the RMS errors of the minimum and maximum temperature forecasts. Official forecasts are, presently, only issued to the public for the first 4 days. There is a substantially improved performance (since the 2001 trial) of the Internet Forecasts (relative to the official forecasts) for the first 4 days (readers may care to contrast Fig. 25 with Fig. 4). In 2001, the CSS for the Internet Forecasts for these days averaged 33.7, the official forecasts 42.7, a difference of 9.0. The hope had been that the system's deficiency (in comparison with the official forecasts) could be somewhat reduced in the 2002/2003 trial. For the first 4 days, using the CSS, that deficiency appeared to have been not only reduced, but, in fact, eliminated. In 2002/2003, the CSS for the Internet Forecasts for Day-1 to Day-4 averaged 31.6, while the official forecasts averaged 30.7. That the Internet Forecasts are now displaying similar levels of overall skill to the official forecasts across the first 4 days, is a reflection of the impact of the new knowledge added to the system.

7.2 Replicating the forecast process

That the CSS for the Internet Forecasts and the official forecasts are now almost the same suggests that the goal of replicating the processes involved in deriving the official forecasts might be close to being achieved. However, a skill score that combines all features of a forecast disguises individual components. When one examines the five verification measures separately, it appears that the similar (overall) CCS values across the first 4 days are largely achieved through:

To illustrate the aforementioned statement, the values of each of the five verification measures across the first 4 days (400 forecasts) for the Internet Forecasts (INT), the official forecasts (OFF), and climatology (CLI), are now presented:

7.3 Performance for Day-1

By contrast, the goal of replicating the processes involved appears almost to be achieved at Day-1, the values of the five verification measures suggesting that the Internet Forecasts display an "across the board" level of skill that is very close to that of the official forecasts:

These data yield Combined Skill Scores (CSSs) of 52.1 (INT) and 51.7 (OFF).

7.4 Performance using the perfect prog assumption (Day-0)

Temperature predictions derived under the perfect prog assumption are the same as the Day-1 Internet Forecasts of temperature (RMS Error of 2.01 deg C). But, Internet Forecasts of rainfall derived under the perfect prog assumption are substantially worse than the Day-1 Internet Forecasts of rainfall (for example, QPF RMS Error of 0.954 Ranges versus 0.714 Ranges). The inferiority (in regard to the rainfall forecasts) places a question mark over the validity of using the 700 hPa RH as a proxy for precipitable water through the entire atmosphere. That, by contrast, Day-1 and Day-2 Internet Forecasts of rainfall are much better suggests that, by 24h and 48h, the NWP model may have distributed moisture sufficiently evenly through the atmosphere to render valid (for Day-1 and beyond) the use of the 700 hPa RH as a proxy for precipitable water.

7.5 Note on the 2002/2003 evaluation

The Internet Forecasts for rainfall are for the 24 hour period commencing 9am, as also are the climatology forecasts with which they are compared. By contrast, the corresponding official forecasts are for the 24 hour period commencing midnight. In order to ensure that the verification data might be properly compared, the official rainfall verification statistics are interpolated to obtain measures that (also) correspond to the 24 hour period commencing 9am.

8. Public forecasts out to 7 days?

Briefly diverting from the main aim of this work, the verification data associated with the official forecasts were examined, with a view to exploring whether it might now be time to issue the experimental official (seven day) forecasts to the public. Such examination suggests that, indeed, it might now be time.

In a similar experiment, conducted over 1997-1998, Stern (1999) concluded that "routinely providing or utilising forecasts beyond day 4 would be inappropriate". However, the new data obtained in the 2002/2003 exercise suggest a change to that conclusion.

To illustrate:

It is interesting now examine the Bureau of Meteorology's most recent report into the skill displayed by the various NWP products (Bureau of Meteorology, 2002a). One finds, firstly, that both the MSL Pressure S1 Skill Score, and the corresponding Anomaly-Correlation (Kalnay et al., 1990) for the Australian NWP Global Analysis and Prediction (GASP) Model are below those corresponding to the European Centre for Medium Range Weather Forecasting (ECMWF) Model for all time periods (Fig. 26). Furthermore:

The profound implication of the aforementioned comments is that the Internet Forecasts have the potential to perform even better (than in the current trial), should data from the ECMWF model be utilised as input, rather than data from the GASP model.

Furthermore, an area of additional knowledge that could be readily added to the structure of the system is that of using identification of the synoptic type as a trigger to vary how the system arrives at its forecasts. To illustrate, it seems inconceivable that, for every synoptic type, the "best" predictors for maximum temperature are:

Using synoptic type as a trigger to vary how the system arrives at its forecasts, should lead to immediate improvement in the accuracy of its output.

There is, also, the possibility that the Internet Forecasts have the potential to perform even better for rural locations, relative to current official forecasts, than they do for Melbourne. This is because forecasters devote more attention to the preparation of forecasts for the city of Melbourne, than they do to the preparation of forecasts for remote rural localities (refer to Stern, 1998 & 1999). By contrast, an automated system would be expected to treat each location equally.

To illustrate this point, Fig. 27 measures the extent that official maximum temperature forecasts for seventeen Victorian centres, issued several days ahead, improve upon objective guidance. At first sight, Fig. 27 appears to suggest that, overall, whereas human forecasters are capable of significantly improving upon computer generated guidance for short-term predictions, that capability is much reduced for long-term predictions. This reduction in performance may be attributed to less attention being able to be directed towards the lower priority long-term predictions.

There is, however, more to this story. Fig. 28 indicates that where forecasters focus on a particular location, as they might be expected to do for Melbourne, the State capital, that capability is somewhat preserved.

9. A Forecasting Research Tool

The system for generating the Internet Forecasts is ideally placed to be a research tool to enhance understanding, not only of the forecasting process, but of weather itself. As Stern (1996 & 1999b) writes:

'Ramage (1993) proposes that the "scientific impact of weather forecasting on understanding weather could be enormous ... (but) we have allowed meteorology to split into two sub-disciplines separated by the activity that could have provided the scientific cement - forecasting". He suggests that poor forecasts are a consequence of poor understanding and presents an iterative procedure whereby understanding (that is, accepted theory) can be modified and then lead to forecast improvement. This is diametrically opposed to what he terms "the bureaucratic viewpoint ... (that) understanding already exists". With that mindset, deficiencies in the current pathway are seen as temporary - more powerful computers being regarded as means (in the future) to overcome problems in the numerical processing; more data being regarded as a means (in the future) to overcome inaccurate specification of the initial state. An iterative procedure is proposed whereby forecast performances, and subsequent case studies, are used in a systematic manner as vehicles to evaluate and modify theory which then feeds directly back into the forecast process.'

In order to contemplate how some views have changed, and some views have not, the aforementioned words may be compared, and contrasted, with the following words, written by the author some 20 years earlier (Stern, 1980):

'The analogue statistics model allows for a considerable degree of man-machine interaction. For example, the operator of the model determines whether analogues to the current situation or to a prognosis of the morrow's situation are withdrawn, the number of such analogues, and the weights to be given to each of the variables used to specify the situation. Furthermore, the operator determines which terms are included in the regression equations which are established for the prediction of a number of weather elements.

It is felt that the role of the operator in the man-machine interaction step should be systematic not arbitrary. It is therefore considered that the weights given to the variables used in the analogue retrieval step, the terms included during the statistical analysis step, etc. should be defined by a set of conditions related to the data. as experience with the use of the model is gained this set of conditions may be modified in order to improve on its overall performance. Indeed, on the basis of the results of the spring-1978 trial, the set of conditions were altered. The model's overall performance during the re-run was slightly better than that of the official Bureau product.'

The analogue statistics model was subjected to further testing (Stern and Dahni, 1981; Stern and Dahni, 1982), some modification, and then implemented operationally in the Bureau of Meteorology's Victorian Regional Office (Dahni et al., 1984a&b; Dahni, 1988). Subsequently, it was generalised, and then implemented in other Bureau Regional Offices (Dahni and Stern, 1995).

9.1 Illustration 1

A simple illustration is now presented of how the knowledge-based system for generating Internet Forecasts might be employed as a research tool to enhance understanding. Fig. 29 depicts the MSL pressure (MSLP) distribution for Synoptic Type 41, a pattern associated with strong cyclonic SSW flow. Fig. 30 plots the annual march of Probability of Precipitation forecasts for strong cyclonic SSW flow, with 850 hPa temperatures of 0 deg C, and with 700 hPa relative humidities of 30% and 70%. Fig. 31 plots the annual march of QPFs for the same synoptic situation, also with 850 hPa temperatures of 0 deg C, and with 700 hPa relative humidities of 30% and 70%. One might reflect upon the reasons as to why there is such a marked difference between the winter and summer responses to the same underlying situation.

An illustration is now presented of how one may proceed from the enhanced understanding achieved via the preceding analysis, to increase the potential accuracy of the forecasts generated. A study of Synoptic Type 41 cases suggests that the strength of the SSW flow between SE Australia and over waters to the west of that region might be significant. Regression analysis confirms that this is, indeed, the case. Adding the terms cyclonicity, (cyclonicity)*(sine day of the year), and (cyclonicity)*(cosine day of the year), to the other 8 predictors, cyclonicity proves to be the most significant of the (now 11) predictors. Fig. 32 plots the annual march of Probability of Precipitation forecasts for strong cyclonic SSW flow with the new equation (850 hPa temperatures of 0 deg C, and 700 hPa relative humidities of 30%). It reveals just how important cyclonicity is in assessing the likelihood of precipitation - demonstrating that, provided there is a deep SSW flow over waters to the west of SE Australia, even in the summer and with low relative humidities, the probability of precipitation is quite high.

9.2 Illustration 2

A second simple illustration is now presented of how the system for generating Internet Forecasts might be employed as a research tool to enhance understanding. Fig. 33 depicts the MSL pressure (MSLP) distribution for Synoptic Type 32, a pattern associated with moderate anticyclonic ENE flow. Fig. 34 plots the annual march of Maximum Temperature forecasts for moderate anticyclonic ENE flow, with 850 hPa temperatures of 7 deg C, with 700 hPa relative humidities of 30%, and with a previous day's maximum temperature of 19 deg C. One might reflect upon the possibility that, particularly during the summer half of the year, the strength of the cooling sea breeze might be governed by the strength of the northerly component of the flow across Victoria for this synoptic pattern. A study of Synoptic Type 32 cases suggests that the strength of the northerly component of the flow across Victoria might be significant. Regression analysis confirms that this is, indeed, the case. Adding the terms north, (north)*(sine day of the year), and (north)*(cosine day of the year), to the other 8 predictors, where

north=MSLP Gabo - MSLP Gambier

north proves to be the second most significant of the (now 11) predictors. Fig. 35 plots the annual march of Maximum Temperature forecasts for moderate anticyclonic ENE flow, with different strengths in the northerly component of the flow across Victoria, once again with 850 hPa temperatures of 7 deg C, with 700 hPa relative humidities of 30%, and with a previous day's maximum temperature of 19 deg C. It reveals just how important north is in assessing the maximum - demonstrating that, in the summer half of the year, a modest northerly component in the flow across Victoria may lead to maximum temperatures about 5 deg C higher than otherwise.

10. Potential

A worthwhile step has been made towards achieving computer replication of the processes involved in how humans derive a weather forecast. Once achieved, such replication allows for systematic incorporation of new procedures that may lead to a quantum leap in the accuracy of the forecasts.

The verification data suggests that perhaps, just perhaps, this goal might already have been attained in regard to replicating the maximum temperature forecasting processes.

Given the exceptionally poor performance of the Internet Forecasts specifically at predicting precipitation at Day-3 and beyond, and also at predicting temperature at Day-6, there is potential (in the short-term) for further overall improvement (of the Internet Forecasts relative to the official forecasts) once the cause of those particular deficiencies are diagnosed and addressed.

Finally, as discussed at the 18th IIPS Conference, the automated nature of the system lends itself:

Regarding the provision of forecasts for numerous localities, although the system, in its present form, has been drastically "scaled back", with there being now only three localities with forecast information aside from Melbourne, the potential for adding other localities is (almost) unlimited.

An illustration is now presented of the system's capability in this regard. Fig. 38 depicts the MSL pressure (MSLP) distribution for Synoptic Type 14, a pattern associated with weak anticyclonic ESE flow. Fig. 39 plots the annual march of Maximum Temperature forecasts for weak anticyclonic ESE flow, with 850 hPa temperatures of 10 deg C, with 700 hPa relative humidities of 30%, and with a previous day's maximum temperature of 19 deg C.

The annual march depicts how, under these circumstances, the (relatively) inland suburb of Watsonia is somewhat warmer during the summer half of the year (by about 1 or 2 deg C) - when the sea breeze has a stronger impact at Melbourne. However, Watsonia is somewhat cooler than Melbourne in winter (by about 1 or 2 deg C).

The annual march also depicts how, under these circumstances, the northern plains of Victoria are warmer than the alpine regions to a greater extent in spring and summer (by about 14 to 15 deg C) - when the lapse rate is greater. By contrast, in autumn and winter, Mildura is warmer than Mt Hotham only by about 11 deg C.

Regarding Indigenous languages, refer here also to the Indigenous Weather Knowledge WebSite (Bureau of Meteorology, 2002b), the associated Media Release (Bureau of Meteorology, 2002c), the associated Newspaper Article (Cauchi, 2003), and also to our Service Charter (Bureau of Meteorology, 2002d).

Regarding Chinese, Gillies (2001) notes that "55% of the web is in languages other than English", and it is becoming increasingly important to cater for these other languages. Chinese is the world's most widely spoken language, and Whittaker (2001) contemplates the possibility of it becoming "the common language of the Web".

The importance of providing multi-lingual weather forecasting and warning services is starkly underlined in a recent report about the fury wrought by Tropical Cyclone Amy as it swept through the South Pacific (Harvey, 2003). The report, which appeared in The Australian newspaper of January 16, 2003, notes that "Waves of up to 30m high swamped entire villages ... (that) eight Fijians are feared dead ... (and that) weather alerts during the cyclone were only broadcast in Fijian, so Hindi and English speakers had no information (my emphasis)."

Earlier in this paper, Brooks (1995), has been quoted as suggesting that technology will create the opportunity for greater concentration on severe events. To this end, the system described herein has been endowed with a facility to "alert" forecasters of the possibility of unusual events such as severe storms, frosts and strong winds (refer to Fig. 40).

11. Concluding remarks

The 100-day verification trial suggests that substantial progress, albeit uneven, towards achieving computer replication of the manual forecast process, has been made. Specifically, the impact of the "locked in" improvements upon the skill displayed by the system has been considerable, particularly for Day-1 forecasts. Nevertheless, the performance of the system over the 100-day trial is so impressive, that one must entertain the possibility that it was simply a "fluke". For this reason, further testing needs to be undertaken. Also, an investigation of the likely cause of the deficient performance at forecasting precipitation beyond Day-1 needs to be carried out, with the consequent incorporation of any new knowledge that such an investigation suggests. Should the further testing confirm the 100-day trial results, we could see ourselves at the "dawn" of the operational implementation of Ramage's approach to weather forecasting which, through computer replication of the manual process, would allow for:

Confirmation of the trial results may have a ramification that extends far beyond providing forecasts simply for Melbourne, Victoria, or even Australia. Suppose one employs a global NWP model to input all of the data required by a system such as the one described here. The ramification is that one might then be able to automatically generate and disseminate, via the internet, forecasts (of a comparable quality to what is now produced manually) for anywhere in the world. Furthermore, this (all-encompassing) generation and dissemination could be carried out at a single location.

12. References

Bureau of Meteorology, 1977. Recording and encoding weather observations. Bur. Met., Australia.

Bureau of Meteorology, 2002a. Analysis and prediction. Quarterly summary. July-September 2002. Bur. Met., Australia.

Bureau of Meteorology, 2002b. Indigenous Weather Knowledge. Bur. Met., Australia.
Refer to: http://www.bom.gov.au/iwk .

Bureau of Meteorology, 2002c. Indigenous Weather Knowledge Website launched today. Bur. Met., Australia.
Refer to: http://www.bom.gov.au/announcements/media_releases/ho/20021220.shtml .

Bureau of Meteorology, 2002d. Service charter for the community. Bur. Met., Australia.
Refer to: http://www.bom.gov.au/inside/services_policy/serchart.shtml .

Brier, G. W., 1950: Verification of forecasts expressed in terms of probabilities. Mon. Wea. Rev., 78:1-3.

Brooks, H., 1995: Human forecasters and technology. Obtained (September 1997) from http://www.nssl.uoknor.edu/~brooks/ and as quoted by Stern (1998).

Brunet, N., Verret, R. and Yacowar, N., 1988: An objective comparison of model output statistics and "perfect prog" systems in producing numerical weather element forecasts. Wea. Forecasting, 3, 273-283.

Cauchi, S., 2003. Bureau wants to tap 50,000 years of weather lore. The Age. 7 January, 2003.
Extract available at: http://www.weather-climate.com/iwk .

Dahni, R. R., 1988: The development of an operational analogue statistics model to produce weather forecast guidance. Ph. D. Thesis, Depart. of Meteorology, University of Melbourne.

Dahni, R. R., 2003: An automated synoptic typing system using archived and real-time NWP model output. 19th Conference on Interactive Information Processing Systems, Amer. Meteor. Soc., Long Beach, California.

Dahni, R. R., de la Lande, J. and Stern, H., 1984a: Testing of an operational statistical forecast guidance system. Aust. Met. Mag., 32, 105-106.

Dahni, R. R., Jasper, J. D., and Stern, H. 1984b: The application of a Markov model to the short-term forecasting of weather state, particularly precipitation. Extended Abstracts Part 2, Conference on Australian Rainfall Variability, Arkaroola, August 1984,Australian Academy of Science and Bureau of Meteorology, 107-110.

Dahni, R. R. and Stern, H., 1995: The development of a generalised UNIX version of the Victorian Regional Office’s operational analogue statistics model. BMRC Res. Rep. 47, Bur. Met., Australia.

Dawkins, S. S. and Stern, H., 2003: Trends and volatility in the accuracy of temperature forecasts. Submitted to the 7th International Conference on Southern Hemisphere Meteorology and Oceanography, Amer. Meteor. Soc., Wellington, New Zealand.

Dobryshman, E. M., 1972: Review of forecast verification techniques. WMO Technical Note, No. 120.

Gillies, M., 2001: Tongues tied. The Australian, 6 August, 2001.

Godau, J., Shepherdley, C., Tan, E., and Silverley, G. 2001: Autotaf - The design, development and implementation of an automated tool for the production of terminal aerodrome forecasts. 17th Conf. on Interactive Information and Processing Systems for Meteorology, Oceanography and Hydrology, Amer. Meteor. Soc., Albuquerque, New Mexico.

Harvey, C. 2003: Fijians killed by massive waves. The Australian, 16 January, 2003.

Kalnay, E., Kanamitsu, M., and Baker, W. E. 1990: Global numerical weather prediction at the National Meteorological Center. Bull. Amer. Meteor. Soc., 71, 1410-1428.

Marzban, C., 2003: Neural networks for non-linear post-processing of model output. 3rd Conference on Artificial Intelligence Applications to Environmental Science, Amer. Meteor. Soc., Long Beach, California.

Noone, D. and Stern, H., 1995: Verification of rainfall forecasts from the Australian Bureau of Meteorology's Global Assimilation and Prognosis System (GASP). Aust. Meteor. Mag.,44:275-286.

Ramage, C. S., 1993: Forecasting in meteorology. Bull. Amer. Meteor. Soc., 74, 1863-1871.

Ryan, C. J., Jha, A., and Joshi, S. 2003: Mentor - a performance support system for forecasters. 19th Conference on Interactive Information and Processing Systems, Amer. Meteor. Soc., Long Beach, California.

Stanski, H., Wilson, L.J., and Burrows, W. 1989: Survey of common verification methods in meteorology. WMO World Weather Watch Tech. Rep. 8, 114 pp.

Stern, H., 1980a: The development of an automated system of forecasting guidance using analogue retrieval techniques. M. Sc. Thesis, Depart. of Meteorology, University of Melbourne (subsequently published by Bur. Met., Australia, in 1985, as Meteorological Study 35).

Stern, H., 1980b: A system for automated forecasting guidance. Aust. Meteor. Mag, 28:141-154.

Stern, H., 1994a: Enveloping predictor parameters in a regression equation - an alternative to model output statistics (MOS). Abstracts, 12th Australian Statistical Society Conf., Monash University, Melbourne, Australia, 4-8 July, 1994.

Stern, H., 1994b: Enveloping predictor parameters in a regression equation - an alternative to model output statistics (MOS). 10th Conference on Numerical Weather Prediction, Amer. Meteor. Soc., Portland, Oregon.

Stern, H., 1996: Statistically based weather forecast guidance. Ph. D. Thesis, School of Earth Sciences, University of Melbourne (subsequently published by Bureau of Meteorology, Australia, in 1999, as Meteorological Study 43).

Stern, H., 1997: A statistically based approach to nowcasting (0-24 hours) on the mesoscale. AIFS Nowcasting Workshop, 24-26 June 1997, Bureau of Meteorology, Head Office, Melbourne.

Stern, H., 1998: An experiment to establish the limits of our predictive capability. 14th Conference on Probability and Statistics /16th Conference on Weather Forecasting and Analysis, Amer. Meteor. Soc., Phoenix, Arizona.

Stern, H., 1999a: An experiment to establish the limits of our predictive capability of weather elements for Melbourne. Australian Meteorological Magazine, 48, 159-167.

Stern, H., 1999b: Statistically based weather forecast guidance. Meteorological Study 43, Bureau of Meteorology, Australia.

Stern, H., 2002a: A knowledge-based system to generate internet weather forecasts. 18th Conference on Interactive Information and Processing Systems, Amer. Meteor. Soc., Orlando, Florida.
Refer to: http://www.weather-climate.com/fc.html .

Stern, H., 2002b: A knowledge-based system to generate internet weather forecasts. Powerpoint Presentation to 18th Conference on Interactive Information and Processing Systems, Amer. Meteor. Soc., Orlando, Florida.
Refer to: http://www.weather-climate.com/internetforecasts.ppt .

Stern, H., 2003a: Progress on a knowledge-based internet weather forecasting system. 19th Conference on Interactive Information and Processing Systems, Amer. Meteor. Soc., Long Beach, California.
Refer to: http://www.weather-climate.com/internetforecasts.html .

Stern, H., 2003b: Progress on a knowledge-based internet weather forecasting system. Powerpoint Presentation to 19th Conference on Interactive Information and Processing Systems, Amer. Meteor. Soc., Long Beach, California.
Refer to: http://www.weather-climate.com/InternetForecastsAMS2003.ppt .

Stern, H. and Dahni, R., 1981: Further testing of Stern's (1980) system for automated forecasting guidance. Aust. Meteor. Mag., 29:69-70.

Stern, H. and Dahni, R., 1982: An automated system for forecast guidance. Abstracts, 4th Conf. on Science Technology, Melbourne, August 1982 (ANZAAS-AIST).

Stern, H., Jasper, J. D., Dahni, R. R., and de la Lande, J., 1984a: Frequency distribution of weather phenomena, particularly precipitation type, from data associated with synoptic analogues. Extended Abstracts Part 1, Conference on Australian Rainfall Variability, Arkaroola, August 1984,Australian Academy of Science and Bureau of Meteorology, 31-34.

Stern, H., Jasper, J. D., and Wilson, K. J., 1984b: Synoptic-scale forcing of "cool-change" events and associated meso-scale temperature variations. Abstracts, Inernational Conference on Meso-scale Meteorology, Melbourne, Australia, February, 1984 (Bureau of Meteorology, Australia), p 7.

Stern, H., de la Lande, J., Dahni, R. R., Jasper, J. D., and Wilson, K., 1987: Meso-scale forecast guidance at the Victorian Regional Office. Operational forecasting. Papers from the Senior Forecasters' Conference June 1985. Training Note No. 2. Bureau of Meteorology,Australia, pp 175-179.

Stern, H. and Parkyn, K., 1999: Predicting the likelihood of fog at Melbourne Airport. 8th Conference on Aviation, Range and Aerospace Meteorology, Amer. Meteor. Soc., Dallas, Texas.

Stern, H. and Parkyn, K., 2001: A web-based Melbourne Airport fog and low cloud forecasting technique. 2nd Conference on Fog and Fog Collection, St John's, Canada.

The Economist, 2003: Why a dumb machine can easily beat a smart human. As quoted by The Australian, 1 February, 2003.

Treloar, A. B. A.. and Stern, H., 1993: A climatology and synoptic classification of Victorian severe thunderstorms. 4th International Conference on Southern Hemisphere Meteorology and Oceanography, Amer. Meteor. Soc., Hobart, Australia.

Whittaker, J., 2001: Speaking in tongues. Australian PC World, August, 98-104.


APPENDIX

1. Postscript

So far, further testing is, indeed, confirming the 100-day trial results. A modest post-trial exercise, involving the derivation of forecasts for Day-1 only for the (now) 130 days ended February 1, 2003, yields:

The additional verification data have resulted in a rise in both CSS values (on account of the well-forecast unusually warm and dry weather during the subsequent 30-day period), but have had only a small impact on the difference between the two values. Regarding the individual components that make up the CSS:

The additional testing not only justifies the use of a module to take into account the "climate change" issue of the upward trend in temperature, but even suggests that a greater "climate-change correction" than was applied here is warranted. After 130 days, the mean bias (forecast-observed) of the Day-1 forecasts of minimum temperature was -0.36 deg C, while the mean bias of the Day-1 forecasts of maximum temperature was -0.32 deg C. Nevertheless, the system appears to have adequately predicted most of the extreme weather events during the evaluation period:

Conversely, the system appears to have been justified on most of the occasions when it predicted extreme weather events:

2. An Alternative Approach

There is another approach to computer replication of the manual forecast process, which may be applied together with, or alternatively to, the aforementioned strategy.

As The Economist (2003) notes (in the context of Deep Blue's success at chess playing), "Rather than emulating the complex thought processes of human players, computers simply resort to mindless number-crunching to decide what move to make. Throw enough microchips at the problem - Deep Blue contained hundreds of specialist chess-analysis chips - and it does indeed become trivial."

This neural networking approach is increasingly being applied to addressing problems of weather forecasting, and Marzban (2003) describes how neural networks "are generally useful because, in addition to being able to approximate a large class of functions, they are less inclined to overfit data than some other nonlinear methods."


FIGURES

...

...

...

Map 1 Map depicting the location of the southeast Australian State of Victoria.

...

...

...

...

...

...

Map 2 Map depicting the location of places referred to in the text.

...

...

...

...

...

...

Fig. 1 Segment of the output of the 18th IIPS Conference version of the system,
illustrating the aviation component.

...

...

...

...

...

...

Fig. 2 Segment of the output of the 18th IIPS Conference version of the system,
illustrating the media graphics capability.

...

...

...

...

...

...

Fig. 3 The web-based product that generates time series of expected weather.

...

...

...

...

...

...

Fig. 4 Illustration of the overall accuracy of the forecasts in 2001.

...

...

...

...

...

...

Fig. 5 An illustration of parameter enveloping.

...

...

...

...

...

...

Fig. 6 Precipitable water/850 hPa temperature relationship.

...

...

...

...

...

...

Fig. 7 Long-term upward trend in Melbourne's minimum temperature
(30-year period ended in indicated year).

...

...

...

...

...

...

Fig. 8 Long-term upward trend in Melbourne's maximum temperature
(30-year period ended in indicated year).

...

...

...

...

...

...

Fig. 9 Accuracy of developmental data forecasts.

...

...

...

...

...

...

Fig. 10 NWP MSL prognosis.

...

...

...

...

...

...

Fig. 11 NWP 700 hPa relative humidity prognosis.

...

...

...

...

...

...

Fig. 12 NWP 850 hPa temperature prognosis.

...

...

...

...

...

...

Fig. 13 The system's output, after inserting data from the NWP prognosis into the form.
(Later, you may experiment with the system, by inserting "your own" data into the "live" form, which appears at the end of the http://weather-climate.com/internetforecasts.html web page.)


...

...

...

...

...

...

Fig. 14 Corresponding official forecast.

...

...

...

...

...

...

Fig. 15 RMS Error of the QPFs.

...

...

...

...

...

...

Fig. 16 Percentage Correct Rain/No Rain Forecasts.

...

...

...

...

...

...

Fig. 17 Brier Score of the PoP Forecasts.

...

...

...

...

...

...

Fig. 18 Reliability Diagram for the Day-1 to Day-6 PoP Forecasts.

...

...

...

...

...

...

Fig. 19 Reliability Diagram for the Day-1 PoP Forecasts.

...

...

...

...

...

...

Fig. 20 ROC Diagram for the Day-1 to Day-6 PoP Forecasts.

...

...

...

...

...

...

Fig. 21 ROC Diagram for the Day-1 PoP Forecasts.

...

...

...

...

...

...

Fig. 22 Illustration of the accuracy of the minimum temperature forecasts.

...

...

...

...

...

...

Fig. 23 Illustration of the accuracy of the maximum temperature forecasts.

...

...

...

...

...

...

Fig. 24 Illustration of the overall accuracy of the temperature forecasts.

...

...

...

...

...

...

Fig. 25 Illustration of the overall accuracy of the forecasts.

...

...

...

...

...

...

Fig. 26 Anomaly Correlation for ECMWF and GASP models (compare with Fig. 25).

...

...

...

...

...

...

Fig. 27 The extent that official maximum temperature forecasts (squares) for seventeen Victorian centres, issued several days ahead, improved upon the guidance (diamonds) - from Stern (1998).

...

...

...

...

...

...

Fig. 28 The extent that official maximum temperature forecasts (squares) for the city of Melbourne alone, issued several days ahead, improved upon the guidance (diamonds) - from Stern (1998).

...

...

...

...

...

...

Fig. 29 Synoptic Type 41 (refer to Dahni, 2003).

...

...

...

...

...

...

Fig. 30 The Annual March of the PoPs.

...

...

...

...

...

...

Fig. 31 The Annual March of the QPFs.

...

...

...

...

...

...

Fig. 32 The Annual March of the PoPs should one add cyclonicity to the predictors.

...

...

...

...

...

...

Fig. 33 Synoptic Type 32 (refer to Dahni, 2003).

...

...

...

...

...

...

Fig. 34 The Annual March of the Maximum Temperature Forecasts.

...

...

...

...

...

...

Fig. 35 The Annual March of the Maximum Temperature Forecasts should one add
north (MSLP Gabo-Gambier) to the predictors.

...

...

...

...

...

...

Fig. 36 Segment of the output of the 18th IIPS Conference version of the system,
illustrating the Chinese component.

...

...

...

...

...

...

Fig. 37 Segment of the output of the 19th IIPS Conference version of the system,
illustrating the Chinese component.

...

...

...

...

...

...

Fig. 38 Synoptic Type 14 (refer to Dahni, 2003).

...

...

...

...

...

...

Fig. 39 The Annual March of the Maximum Temperature Forecasts for Synoptic Type 14.

...

...

...

...

...

...

Fig. 40 Illustration of the system's alerting capability.

...

...

...


APPLYING THE SYSTEM
(experiment with the system by entering data in the form below):

Enter Date (for which forecast applies):
1) Date (1,2...31): 2) Month (1,2...12): 3) Year (2002, 2003...):
Days from NWP Base Analysis:
4) Days Ahead (1,2,3,4,5,6):
Enter Forecast 9am MSL Pressure Data (from NWP model):
5) Melbourne: 6) Hay: 7) Smithton:
8) Gabo Is: 9) Mt Gambier: 10) Forrest:
Enter Melbourne and 9am NWP Model Data:
11) Max on day before
forecast day (-24h):
12) 700 hPa % RH
on forecast day:
13) 700 hPa % RH
on day after (+24h):
14) 850 hPa Temp
on forecast day:
15) 850 hPa Temp
on day after (+24h):
Synoptic Information:
16) Direction: 17) Curvature: 18) Speed:
Synoptic Type (click below to view):
19) Type:
Forecast Rain (24h from 9am):
24) PoP (%): 25) QPF (mm):
Forecast Temperature:
26) Min (deg C)
during preceding night:
27) Max (deg C)
during current day:
Additional Localities:
28) Max (deg C)
at Mildura:
29) Max (deg C)
at Mt Hotham:
30) Max (deg C)
at Watsonia:
WEATHER FORECAST:
WIND FORECAST:
LINK TO FORECAST
IN CHINESE:
(click below to view)

LINK TO FORECAST IN CHINESE: Note: After clicking on the number corresponding to the appropriate forecast in Chinese, a new window, with a forecast in Chinese, will appear. Exiting from the new window will return you to the original window. 1 2 3 4 5 6 7 8 9 10 11 12 13 14

VIEW SYNOPTIC TYPE: Note: After clicking on the number corresponding to the appropriate synoptic type, a new window, with an analysis, will appear. Exiting from the new window will return you to the original window. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50


© Copyright Commonwealth of Australia 2002, Bureau of Meteorology (ABN 92 637 533 532)
Please note the Copyright Notice and Disclaimer relating to the use of the information on this site and our site Privacy Policy.
Users of these web pages are deemed to have read and accepted the conditions described in the Copyright Notice and Disclaimer.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.