Bayesian model averaging (BMA) is a popular method using the advantages of forecast ensemble to enhance the reliability and accuracy of predictions. The inherent assumptions of the classical BMA has led to different variants. However, there is not a comprehensive examination of how these solutions improve the original BMA in the context of streamflow simulation. In this study, a scenario-based analysis was conducted for assessment of various modifications and how they affect BMA results. The evaluated modifications included using various streamflow ensembles, data transformation procedures, distribution types, standard deviation forms, and optimization methods. We applied the proposed analysis in two data-poor watersheds located in northern Ontario, Canada. The results indicate that using more representative distribution types do not significantly improve BMA-derived results, while the positive effect of implementing non-constant variance on BMA probabilistic performance cannot be ignored. Also, higher reliability was obtained by applying a data transformation procedure; however, it can reduce the results’ sharpness significantly. Moreover, although considering many streamflow simulations as ensemble members does not always enhance BMA results, using different forcing precipitation scenarios besides multi-models led to better BMA-based probabilistic simulations in data-poor watersheds. Also, the reliability of the expectation-maximization algorithm in estimating BMA parameters was confirmed.
Probabilistic streamflow forecasting is of increasing interest in various fields of water resources management from real-time flood forecasting to long-term management of water systems. Accurate and reliable short-to-medium-range streamflow forecasts, with lead-times ranging from hours to days, can play an important role in flood control, mitigation, and early warning systems (Bravo et al., 2009;Thiemig et al., 2015). Unlike deterministic forecasts, which provide a point estimation of the river flow, probabilistic forecasts try to quantitatively assess the inherent uncertainties associated with the streamflow predictions and provide a predictive uncertainty distribution, which is required for reliable and informed decision making (Biondi
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.