Predictive models

We can define predictive models as quantitative mathematical projections that use statistical classifiers to determine the probability of a specific water quality event in the future.

Predictive modelling can also be applied to an unknown water quality event, even after it has occurred.

Developing and using a more quantitative ‘water quality model’ will improve:

current understanding of the system (typically depicted in a conceptual model) at Step 1 of the Water Quality Management Framework
predictions of water quality improvement from management strategies at Step 8 of the framework.

Types of predictive models for water quality

There are almost as many schemes for classifying water quality models as there are models. One useful classification scheme (SKM 2011) describes 3 types of water quality models:

catchment models — derive flows from rainfall runoff and simulate associated pollutant loads
in-stream models — simulate hydrodynamic behaviour of flows and in-stream water quality processes
ecological response models — simulate the ecosystem response to stressors, such as flow and water quality.

These types of models need to be highly interdependent (Webster et al. 2008). Bayesian networks and other statistical interaction models have been used as a high-level integration framework.

Catchment models

Catchment models are highly developed and widely used in Australia.

Australia's national hydrological modelling platform, eWater Source, has been used in many locations throughout the country.

In-stream (or receiving waters) models

In Australia, there are scattered occurrences of the use of in-stream models for rivers, with no consistency in application.

The widely used Integrated Quantity and Quality Model (IQQM) has been promoted as having water quality modelling capabilities but the little development in this area has focused on ‘conservative’ water quality constituents, such as salinity.

Hydrodynamic and water quality models have been applied many times in Australia’s estuarine and coastal waters. Refer to CSIRO’s eReef suite of models.

Ideally, in-stream models would be coupled with catchment models. Most of the applications to date simply use the outputs from the catchment models as inputs to the in-stream models.

Ecological response models

Ecological response models are only useful for management goals relevant to the community value of ecosystem protection. In Australia, they have been developed to understand the magnitude of ecological change in the Murray–Darling Basin under multiple scenarios.

Eco Modeller is an example of a tool that can be used to build, store and run quantitative models of ecological responses to physical and biological factors.

In the Water Quality Guidelines, our approach to deriving toxicant guideline values could be considered a type of ecological response model.

Guidance on application of water quality models

Formulating the conceptual model at Step 1 of the Water Quality Management Framework is the precursor to development of more detailed quantitative prediction models. The more complex the conceptual modelling, the more detailed a predictive model may be.

Determination of optimal model complexity is important. Sometimes more complex models can be less successful as a predictive tool than simpler models due to lack of good quality calibration data.

Predictive models are very useful at Steps 8 and 9 of the framework, where various management strategies are evaluated. Predictive modelling can help to decide the most effective strategy, as well as the feasibility of achieving the required water quality objectives.

Outputs of a predictive model need to be regularly tested and re-calibrated with improved understanding, and the model adjusted via a well-designed monitoring and evaluation program.

Models also play a part in the assessment of achieving water/sediment quality objectives. For example the Paddock to Reef Integrated Monitoring, Modelling and Reporting Program in the Great Barrier Reef.

Model uncertainty is an important issue that is often downplayed.

Many models suffer from a lack of reliable data for calibration and verification purposes. In this case, it becomes particularly important to clearly identify model assumptions and the uncertainty created, to avoid creating a false impression of accuracy. The uncertainty in model output should be an important factor when making decisions.

See also: