Practical guidance on representing uncertainty in hydrological predictions
Understanding the uncertainty in hydrological model predictions is critically important for risk assessment and managing water resources. The appropriate statistical representation of residual errors (i.e. differences between simulated and observed flows) is essential for accurate and reliable probabilistic predictions of streamflow. Residual errors of hydrological predictions are often (i) heteroscedastic, with magnitude of error increasing with magnitude of runoff, and (ii) persistent, with errors exhibiting temporal autocorrelation. In this talk we will illustrate why heteroscedasticity and persistence are important, and how different representations of these characteristics can influence predictive performance. Importantly, for hydrological practitioners, our findings will provide practical guidance on the selection of approaches for modelling heteroscedasticity and persistence. This will enhance their ability to provide hydrological probabilistic predictions with the best reliability and precision for different catchment types.