Advances in improving streamflow predictions, with applications in forecasting
Seminar presented at the Bureau of Meteorology, 29th March 2017Improving streamflow predictions including uncertainty quantification is of major interest to the environmental research community and operational agencies including the Australian Bureau of Meteorology. Reliable quantification of predictive uncertainty is important for risk assessment and managing water resources. The University of Adelaide (UoA), the University of Newcastle (UoN) and the Bureau have collaborated closely since
... achieving a series of advances for the Bureau's seasonal streamflow forecasting (SSF) system developed by Extended Hydrological Prediction Team. This talk focuses on recent advances in quantifying streamflow predictive uncertainty using residual error models for daily rainfall-runoff models at the scale of interest (eg, daily or monthly). To achieve reliable and precise probabilistic predictions across a range of time scales, residual error models must capture the strong heteroscedasticity and temporal persistence of streamflow predictive errors. The talk presents the outcomes of a comprehensive comparison of multiple error models over 23 perennial and ephemeral catchments using multiple performance metrics, and interprets the empirical results using theoretical insights. The study provides practical guidance for hydrological practitioners on how to improve the reliability and precision of probabilistic streamflow predictions over a range of catchment types (perennial/ephemeral). Applying the outcomes of this study in a seasonal forecasting context using the Bureau's SSF system over a wide range of 300+ catchments confirms substantial reductions in streamflow forecast uncertainty compared to alternative approaches. In addition, the seminar provides an overview of broader advances in hydrological modelling by UoA/UoN, including flexible hydrological models in lumped and distributed applications, model calibration (optimisation) and spatial rainfall modelling.