What are Ensembles?

In the context of the DSS, an ensemble is a set of time series having some statistical similarity that are handled together. They are generated through some probabilistic techniques. Basically it is a group of possible behaviors of the system that reflect the inherent or external uncertainties due to initial conditions, parameters, or forcing of the system.


What are the uses of ensembles?

Ensembles are used extensively in climate science, weather prediction, hydrological forecasting, climate change studies, etc. Any member of the ensemble is sometimes called a trace. Individual traces are not usually important; it is the statistics of these traces that is important. This is one aspect of the “handled together” attribute of an ensemble. This togetherness also implies that the model will automatically run for all ensemble members and produces the output as an ensemble as well. However, in some cases, the output ensemble members can be generated separately and then grouped to form the ensemble.

When a model is forced with an ensemble input, the results are basically an ensemble of outputs which are analyzed statistically, i.e. statistics like the ensemble mean or median gives us an idea of the average system behavior while the range or standard deviation across the ensemble give us the uncertainty range in the output. The main reason to use ensembles is that we are not certain about the behavior of the system under consideration.


How are ensembles generated?

Ensembles can be generated in many different ways:

• Using different initial conditions, e.g. different initial lake water levels, one can generate an ensemble of lake outflows; different initial weather conditions can be used to generate an ensemble of different future weather predictions; etc.

• Using different model parameters, or even parameterization (i.e. process representation), one can generate an ensemble of model results, e.g. different hydrological parameters that produce similar performance in terms of calibration criteria (Bias, R2, etc.) can be used to generate an ensemble of catchment runoffs. One can use several calibrated hydrological models to generate an ensemble across models which will vary due to the different process representations in each model (e.g. single soil layer rainfall-runoff model vs. a more complex 4 layer model that are both calibrated to the same data will perform differently and they can be used to generate an ensemble)

• Using different forcing, e.g. an ensemble of climate scenarios can be used to force a model to calculate the impacts of these scenarios.

• Mixing more than one aspect of the above when generating the ensemble input leads to an ensemble of ensembles, which is sometime termed, a grand ensemble. For example, if one runs a weather generator with different initial conditions and different model parameters, the result will be a grand ensemble reflecting both factors.