Who will use Seasonal Streamflow Forecasts?

Seasonal Streamflow Forecasts have the potential to help water managers and users make better-informed decisions on:

• seasonal water allocation outlooks
• reservoir operations
• environmental flow management
• water markets
• drought response strategies, including setting supply restrictions.
How do I interpret the historical reference and forecast boxplot product?

This product uses boxplots to represent the forecast and historical reference probability distributions. The forecast probability distribution is shown on the right (in orange) and is constructed from forecast ensemble members generated by a model, with each member having the same probability of occurring. The historical probability distribution is shown on the left (in blue) and is a probabilistic representation of the historical data.

The median, whiskers and box (IQR) of both the forecast and historical reference boxplots can be compared. For example, in the figure below, the median of the forecast is slightly higher than the median of the historical reference, indicating the likelihood of higher than median streamflows. Also, the forecast boxplot shows that the range of streamflow volumes, as represented by the box and the span of the whiskers, is narrower than the historical reference. This is useful information as it gives decision makers an idea of how much confidence can be placed in the forecasts compared with using historical data.

If no forecast could be issued for a particular site (through very low model skill or missing observed data), only one boxplot will be shown for the historical reference as shown below.

How do I interpret the tercile forecast?

The percentage of the forecast in each of the terciles that are determined from the distribution of historical reference data is displayed in a pie chart (shown below). Each of the terciles are labelled so forecasts with the greatest probability of being in the lower third of the distribution are defined as low flow, those in the middle third of the distribution are defined as near median flow and those in the upper third of the distribution are defined as high flow.

What is a boxplot?

A boxplot is also known as a box-and-whisker diagram. The box in the boxplot extends from the 25th percentile to the 75th percentile, with a line at the median (50th percentile), shown in the figure below. The box is known as the Interquartile Range and is a measure of statistical spread for the most likely streamflow occurrence - it covers half of the forecast (or historical reference) distribution. Consequently, the probability of the observation lying in the IQR is 50%.

The whiskers extend from the box and are represented by the bottom and top bars (horizontal lines) at the end of the vertical lines. For the forecasts presented here, the value of the top whisker is the 95th percentile, while the value of the bottom whisker is the 5th percentile. The range between the two whiskers encompasses 90% of the forecast (or historical reference) distribution, therefore observed streamflows are expected to occur within this range 9 times out of 10. There will be some instances where observed streamflows are outside of the whisker extents, representing unusually low or high flow rates.

What is the difference between a site-based forecast and a total catchment inflow forecast?

A site-based forecast uses streamflow data recorded at a gauging station. A total catchment inflow forecast uses estimated streamflow data based on knowledge of the measured streamflow, change in storage level, modelling of releases and regulated flow.

Why are the pie chart forecasts on the tercile summary map faded or grey?

The pie charts are displayed differently to indicate the skill scores of the site forecasts for each season. Forecasts with moderate to high skill are represented by a coloured pie chart. The forecasts with low skill for a particular season are displayed as a faded or transparent pie chart on the tercile summary map. The forecast is still provided. Site forecasts with very low skill or skill the same as the historical reference forecasts are replaced with a grey scale pie split into equal thirds representing the historical reference. For these very low skilled site forecasts the historical reference forecasts are used and the BJP forecasts have been removed.

For example, in the tercile summary map below the site forecast for Darbalara is displayed as a coloured pie chart as it has moderate to high skill for that season. The forecast for the Kiewa catchment is displayed as a faded pie chart to represent that it has a low skill score. The very low skill forecast for Dohertys is shown as a grey pie chart.

Why are the forecasts not shown on some of the products?

The forecasts are not shown for sites that have a very low skill score for that season or for which observed data on antecedent conditions are missing . A very low skill score essentially means the forecast will not exceed the skill of a historical reference forecast. For this reason the BJP forecasts have been removed for particular seasons.

Are the Seasonal Streamflow Forecasts related to the Bureau's Seasonal Climate Outlooks?

The statistical technique used for the Seasonal Streamflow Forecasts is different to the dynamical (physics based) technique used for the Climate Outlooks. The climate outlook provides forecasts across the whole of Australia. The Seasonal Streamflow Forecasts are provided at the catchment scale and the predictors and modelling have been developed to maximise the skill at the catchment scale. The seasonal climate outlook predicts the likelihood of rainfall in the coming three months. In future, rainfall forecasts from the Bureau’s climate outlooks will be used to drive an improved technique for producing Seasonal Streamflow Forecasts.

The amount of runoff and streamflow from a catchment is very dependent on the rainfall and catchment conditions such as soil moisture. It is possible to record high flow from a catchment that receives median rainfall when the catchment is saturated. It is also possible to record low flow from a catchment that receives median rainfall if the catchment is very dry. Rainfall intensity and duration can also influence soil moisture and runoff. Some catchments need heavy rainfall to produce runoff. A season with many rain days and low daily rainfall totals may record a high three-monthly rainfall total but still not provide high runoff.

How can high streamflows be recorded when a catchment receives below average rainfall?

Seasonal streamflows at the forecast sites are a function of several factors, including initial catchment conditions, such as soil moisture, antecedent streamflows due to some level of autocorrelation in the streamflow series, and the characteristics of catchment rainfall events during the season. Even though rainfall was below average during April to June 2011, most forecast locations reported above median streamflows during April to July as a result of above to very much above average catchment soil moisture and high antecedent streamflows.

How do I interpret the exceedance probability forecast?

The exceedance probabilities are defined separately for each forecast location, for each three month period. Firstly, the historical streamflow data for a particular site and a particular three month period are arranged in descending order (highest to lowest). Then the rank value is converted to an exceedance probability from 0% to 100%. The forecast exceedance probability indicates the likelihood that a particular streamflow value will be exceeded. These steps are shown in the figure below.

An exceedance probability forecast is shown in the image below for the site Acheron River at Taggerty. As an example, the forecast likelihood of exceeding a streamflow volume of about 100 GL is 80% (indicated by the red arrows). Also, as another example, there is a 20% chance that a streamflow volume of 185 GL will be exceeded (indicated by the black arrows).

How do I interpret the monthly streamflow boxplot?

A boxplot is also known as a box-and-whisker diagram. This boxplot represents the monthly streamflow historical data. The time period of the historical data used is given in the legend e.g. Jan 1900 - Dec 2010. The monthly observed streamflow data for the last 12 months is represented by the pink dots and is joined by a dotted line.

The box extends from the 25th percentile (lower quartile) to the 75th percentile (upper quartile), with a line at the median (50th percentile). Percentiles are values that divide a set of observations into 100 equal parts. The distance of the 75th percentile minus the 25th percentile is the interquartile range (IQR), which is the height of each box. The whiskers extend from the box and are represented by the bottom and top bars at the end of the blue dotted lines. The value of the top whisker is the highest data value still within the 75th percentile value plus 1.5 multiplied by IQR. The value of the bottom whisker is the lowest data value still within the 25th percentile value minus 1.5 multiplied by IQR. The plus (+) symbols represent outliers or extreme values. Outliers are data points that fall outside of the range of whiskers.

How are the skill score categories defined?

The following definitions can be used in relation to the hindcast RMSEP skill score:

• 0 is considered to be a forecast with no skill (equivalent skill to predicting using historical averages or historical reference)
• less than 5 is considered to be a forecast with very low skill
• 5-15 is considered to be a forecast with low skill
• 15-30 is considered to be a forecast with moderate skill
• greater than 30 is considered to be a forecast with high skill

A forecast with zero skill gives no additional information when compared with the historical reference.

How do I interpret the model validation graphs?

In the forecast quantiles and observations versus year plots, the pink dots indicate the observed value, the dark blue lines are the interquartile range, and the light blue lines are the 5th to 95th percentile range. There should not be any obvious trend with time in the relationship between the forecasts and the observed values.

Why are climate indices used to predict streamflow?

The Seasonal Streamflow Forecasts rely on a modelling approach that was developed by the Commonwealth Scientific and Industrial Research Organisation (CSIRO) under the Water Information Research and Development Alliance (WIRADA). This research describes how climatic conditions influence streamflow forecasts. The following paragraphs are edited excerpts from (Robertson, D.E. and Q.J. Wang 2009).

There are two main sources of predictability in Australian streamflows; initial catchment conditions and climate indices. Strong serial correlations in streamflows arise due to soil and groundwater storages extending the time between the incidence of rainfall and any resulting streamflow. Thus, indicators of initial catchment conditions may be good predictors of future streamflows. Future rainfall and climate also influence future streamflows. Many indices of large-scale climate anomalies, such as the Southern Oscillation Index, Indian Ocean Dipole Mode Index and the Antarctic Oscillation show significant concurrent and lagged correlations with rainfall and streamflows, and therefore may be useful predictors of streamflows too.

A method has been developed to select predictors of streamflows for the BJP modelling approach to seasonal streamflow prediction. The predictor selection method seeks to identify reliable predictors that produce skilful predictions. An important outcome should be that the selected predictors are consistent with our understanding of the physical hydrological and climate systems. The method selects predictors that give the largest improvement in prediction accuracy and are supported by statistical evidence.

The majority of climate indices used in the BJP modelling system are related to sea surface temperature, but there are several related to atmospheric pressure and one related to upper winds.

More climate information is available at Australian climate influences.