A virtual hydrological framework for evaluation of stochastic rainfall models

Stochastic rainfall modelling is a commonly used technique for evaluating the impact of flooding, drought, or climate change in a catchment. While considerable attention has been given to the development of stochastic rainfall models (SRMs), significantly less attention has been paid to developing methods to evaluate their performance. Typical evaluation methods employ a wide range of rainfall statistics. However, they give limited understanding about which rainfall statistical characteristics are most important for reliable streamflow prediction. To address this issue a formal evaluation framework is introduced, with three key features: (i) streamflow-based, to give a direct evaluation of modelled streamflow performance, (ii) virtual, to avoid the issue of confounding errors in hydrological models or data, and (iii) targeted, to isolate the source of errors according to specific sites and seasons. The virtual hydrological evaluation framework uses two types of tests, integrated tests and unit tests, to attribute deficiencies that impact on streamflow to their original source in the SRM according to site and season. The framework is applied to a case study of 22 sites in South Australia with a strong seasonal cycle. In this case study, the framework demonstrated the surprising result that apparently “good” modelled rainfall can produce “poor” streamflow predictions, whilst “poor” modelled rainfall may lead to “good” streamflow predictions. This is due to the representation of highly seasonal catchment processes within the hydrological model that can dampen or amplify rainfall errors when converted to streamflow. The framework identified the importance of rainfall in the “wetting-up” months (months where the rainfall is high but streamflow low) of the annual hydrologic cycle (May and June in this case study) for providing reliable predictions of streamflow over the entire year despite their low monthly flow volume. This insight would not have been found using existing methods and highlights the importance of the virtual hydrological evaluation framework for SRM evaluation.