The Water and the Land (WATL) web site provides rainfall forecast maps generated automatically by weather forecast models. Up to eight models are combined to produce the rainfall forecasts using the method described below in Combining Models. Testing has shown that it easily outperforms a random guess or climatology, to at least five days ahead. This does not mean it is perfect - no forecast system ever will be - but it does mean that people can be confident that on average, the forecast will supply them information to help shape their decisions.

**It is important that you check your local Bureau weather forecasts regularly in conjunction with the rainfall forecasts given in the WATL pages**. Weather forecasters provide supplementary information about the distribution, onset, duration and intensity of the rainfall as well as the likelihood of thunderstorms and other severe weather. Sometimes rainfall patterns can be missed by the combined computer model weather forecasts, for example when showers are just on the coast. At other times the models may overestimate rainfall. Weather forecasters use their experience to determine the most likely outcome from a wide variety of different sources.

The forecast maps have been prepared on a grid roughly 50 km by 50 km. Because rainfall is seldom uniform over such an area (particularly when the rain is falling as showers or thunderstorms, or when local topography is influencing the location of the heavier falls), you should treat the totals as a guide only.

The forecasts are also available from the Bureau's FTP site.

**Schedule**

The rainfall forecast maps offer daily totals, and chance of rain, five days ahead. Maps of 4-day totals for the next 1-4 days, and 5-8 days are also available. The 24-hour rainfall forecasts are updated twice a day at approximately 8 am and 8 pm EST. The 4-day total maps are updated at midnight.

The 24 hour rainfall forecast period ends at 12 UTC (ie.10 pm EST, 9:30 pm CST or 8 pm WST). For example, Tuesday's rainfall will be the amount for rain expected for the period from 10 pm EST Monday evening to 10 pm EST Tuesday evening.

The amount of rain forecast for each day is presented on a national colour coded map, with options to zoom in to state and then district level. Rainfall totals less than 1mm are not shown. The colours represent ranges of forecast rainfall, for example the lightest orange represents a forecast range from 1mm to 5mm.

The chance of rain is the proportion of available models predicting rain at or above the given threshold, expressed as a percentage.

The advantage of using multiple models to determine rainfall is the ability to estimate the chance (otherwise termed probability, or likelihood) of receiving rain. For instance, if seven of the eight models believe at least 10 mm will fall, then the probability of receiving at least 10 mm will be listed as 7/8, or 88%. Likewise, if only one model thinks there will be 10 mm, then the chance of at least 10 mm will be 1/8, or in other words, 13%. No particular model is favoured. Sometimes there may be a "chance of rainfall" when the total expected is less than 1 mm (and is not marked on the rainfall total maps).

Not all models are available out to five days. Hence the 'chance of rainfall' information available for the later days - especially very high values - should be viewed with caution. When only two models are available, if both models agree that there will be rainfall then the chance of rain will be 2/2 or 100%. If there is disagreement the chance will be 1/2 or 50%. In this case there are no intermediate colours in the maps. See the example below. Rarely, when only one model is available, the chance of rain cannot be calculated and a map with "No data available" will be shown.

The maps show the total amount of rain expected for the week ahead in two time blocks, days 1- 4 and days 5 - 8. The forecasts for the first four days are expected to be more accurate than days 5-8. Remember that, as the total forecast rainfall maps are updated at midnight, between 8 pm and midnight there might sometimes be slight differences between the sum of the daily maps and the total rainfall maps.

If a rainfall forecast for any one of the four days is unavailable, then a map with "No data available" will be shown.

The forecast maps are generated automatically by combining the outlooks from a group of Australian and international weather forecast models. Such a combination has been shown to provide a more accurate forecast than using a single model (Ebert 2001). The reason is quite simple. Small errors in the data put into the models - and the different way the models solve all the equations of the atmosphere - means that no two models will forecast the same situation in exactly the same way. In general the different models will tend to "cluster" around the perfect forecast - some a little too wet, some a little too dry. By combining several models we can, on average, get a more accurate forecast than if we just used one model. When using only a single model we will never know in advance if it is making a forecast that is too wet or too dry. However if we use several models, we do know that their combined outlook will tend to be most likely.

Numerous studies have shown that the chaos in the atmosphere, where small and difficult-to-measure changes can result in large scale shifts in weather. The most famous, though unlikely, analogy is the flapping of butterfly wings causing a tropical cyclone. This means that the accuracy of any system declines rapidly the further ahead a forecast is made. Forecasts for later days will almost always have less skill than forecasts for the first day. Despite this decrease in accuracy, forecasts out to five days can provide useful guidance on rainfall trends.

A good probability forecast system should be "reliable". For example, when a forecast suggests there is a 60% chance of rain, we might expect that, on average, 60% of the time rain will indeed fall. We can assess forecast reliability using a "reliability diagram", plotting a graph with the forecast chance of rain on one axis, and observed percentage of time rain occurred, when such a forecast was issued, on the other axis. If the graph lies to the right of the diagonal, it means that the forecasts have tended to be over-confident that the rainfall totals would reach the various thresholds, while to the left of the diagonal indicates the forecasts were not confident enough.

We continually check the performance of the multi-model rainfall forecasts by plotting reliability diagrams over the past 90 days.

Over the development period of the new rainfall forecasting technique, the diagrams have consistently shown reliable forecasts when averaged over these areas. In other words, we can be fairly confident that on average, say, four out of five days will be wet when the forecast is for an 80% chance of rain falling.

Many people are interested in how well the forecast predicted individual rainfall events. While scientists prefer to assess longer time scales and large areas to get the bigger picture - they don't want to miss the wood for the trees - it will always be fascinating to examine individual days of rainfall. Daily verification is available for the last 10 days for an experimental version of the forecasts (PME). Other technical skill scores are also available from the Bureau's Research Centre.

Numerical Weather Prediction (NWP) or computer models included in the calculation of the rainfall totals and the chance of rain are from the:

- Australian Bureau of Meteorology
- US National Oceanographic and Atmospheric Administration

- UK Meteorological Office

- Japanese Meteorological Agency

- European Centre for Medium Range Weather Forecasting

- Meteorological Service of Canada

- German national weather service, Deutscher Wetterdienst

The models vary in the length of time of the forecast and also the grid size, or the distance between points, inside them. Some computer models only provide forecasts out to 3 days ahead while others extend to 10 days ahead.

The models and base times, in UTC, used in our rainfall prediction maps are as follows:

Country or region |
Computer model name |
Grid size in kilometres |
Days forecast |
7:40 am EST issue |
6:30 pm EST issue |
---|---|---|---|---|---|

Australia |
ACCESS-R |
37 |
3 |
12 UTC |
00 UTC |

Australia |
ACCESS-G |
40 |
7 |
12 UTC |
00 UTC |

United Kingdom |
UKGC |
140 |
3 |
12 UTC |
00 UTC |

United States |
USAGFS |
55 |
7.5 |
12 UTC |
00 UTC |

Canada |
CMCGEM |
65 |
6 |
12 UTC |
00 UTC |

Germany |
DWD |
55 |
3 |
12 UTC |
00 UTC |

Europe |
ECSP |
165 |
10 |
12 UTC* |
12 UTC* |

Europe |
ECSP |
55 |
8 |
12 UTC |
00 UTC |

Japan |
JMAGSM |
140 |
8 |
12 UTC |
12 UTC* |

- * previous days model run
- UTC = Coordinated Universal Time

The models are combined to produce rainfall forecasts using a technique known as "probability matched ensemble mean". This technique is used because different models may all forecast rain for a region, but they will tend to place the highest rainfall in slightly different areas. As a result, when models are combined by simple averaging they tend to be accurate in showing where rain will fall, but the actual amounts get "smeared" over a broad area, resulting in an under-forecast of the maximum rainfall. This is particularly noticeable when heavy rain is possible. The PME helps to overcome this problem. While the method sounds complex, it is in fact quite simple.

Suppose at a particular time rain forecasts from 9 weather models are available. For each gridbox (roughly 50 km by 50 km) over Australia, an average value of the 9 different model forecasts is made. These average values are then stored away.

Next the Australia-wide set of the individual model forecasts for every gridbox over Australia is ranked from highest to lowest.

Finally the average values we stored are taken out again, and the grid with the highest average value is assigned the top 9 highest rainfall forecasts from all of Australia. That gridbox is then re-assigned the central value of the top 9. This is then repeated for the grid with the second highest average value - it is re-assigned the value of the median of the model forecasts ranked 10 to 18, and so on.

What this does is to ensure that differences of opinion between the models about the exact location of the rainfall aren't allowed to cloud their collective decision on the amount that will be forecast. It's like saying "well let's just assume that we think just this one region (e.g., the area forecasted to have the highest average rainfall) will get the rain. In that case, what would your model's forecast be?"

Results have shown that by combining the forecasts in this way, high rainfall events are more clearly defined, while low rainfall forecasts show little change in skill. Conversely, as probabilities are calculated using the raw model data, during high rainfall events the maximum predicted rain amount may be high even when the chance of high rainfall is relatively low. It could be said that the forecast totals are a measure of how well the models agree , while the "chance of rainfall" is a representation of how different the location of the rainfall may be.

Ebert, E.E 2001. *Ability of a poor man's ensemble to predict the probability and distribution of precipitation.* Monthly Weather Review, 129, 2461-2480.

© Commonwealth of Australia 2008, Bureau of Meteorology