Practical approaches to produce high-quality probabilistic predictions and improve risk-based design making

Abstract

Probabilistic predictions provide crucial information regarding the uncertainty of hydrological predictions, which are a key input for risk-based decision-making. High-quality probabilistic predictions provide reliable estimates of water resource system risks – avoiding a false sense of security. However, probabilistic predictions are not widely used in hydrological modelling applications because they are perceived to be difficult to construct and interpret. We present a software tool that provides an easy-to-use and simple approach to produce high-quality probabilistic streamflow predictions. The approach integrates the recommendations from multiple research papers over multiple years to provide guidance on selection of robust descriptions of uncertainty (residual error models) for a wide range of hydrological applications. This guidance includes the choice of transformation to handle common features of residual errors (heteroscedasticity, skewness, persistence) and techniques that handles a wide range of common objective functions. A case study illustrating the practical benefits of uncertainty analysis for risk-based decision- making is provided. The case study evaluates fish health in two catchments (Mt. McKenzie and Upper Jacobs) in Barossa Valley, South Australia. The streamflow predictions of environmental flow metrics are combined with a simplified environmental response model to estimate fish health. The outcomes obtained using deterministic streamflow predictions are contrasted to the outcomes obtained from probabilistic predictions. In general, probabilistic predictions provide greater confidence in the predictions of fish health because the uncertainty ranges recognise the differences at the two sites between the quality of hydrological predictions. The uncertainty ranges were generally high, in the range 40-60% (Mt McKenzie) or 4-20% (Upper Jacobs) for predictions of the frequency of years with poor (or worse) fish health. This analysis provides a richer source of information for risk averse decision-makers than the single values provided by deterministic predictions.