An Approach for Improving Prediction in River System Models Using Bayesian Probabilities of Parameter Performance

Abstract:

Achieving predictive success is a major challenge in hydrological modelling. Predictive metrics indicate whether models and parameters are appropriate for impact assessment, design, planning and management, forecasting and underpinning policy. It is often found that very different parameter sets and model structures are equally acceptable system representations (commonly described as equifinality). Furthermore, parameters that produce the best goodness of fit during a calibration period may often yield poor results outside of that period. A calibration method is presented that uses a recursive Bayesian filter to estimate the probability of consistent performance of parameter sets in different sub-periods. The result is a probability distribution for each specified performance interval. This generic method utilises more information within time-series data than what is typically used for calibrations, and could be adopted for different types of time-series modelling applications. Where conventional calibration methods implicitly identify the best performing parameterisations on average, the new method looks at the consistency of performance during sub-periods. The proposed calibration method, therefore, can be used to avoid heavy weighting toward rare periods of good agreement. The method is trialled in a conceptual river system model called the Australian Water Resources Assessments River (AWRA-R) model in the Murray-Darling Basin, Australia. The new method is tested via cross-validation and results are compared to a traditional split-sample calibration/validation to evaluate the new technique’s ability to predict daily streamflow. The results showed that the new calibration method could produce parameterisations that performed better in validation periods than optimum calibration parameter sets. The method shows ability to improve on predictive performance and provide more realistic flux terms compared to traditional split-sample calibration methods.