Applications Determine the Best Model to Predict Flow Duration Curves in Ungauged Basins

Abstract:

Flow duration curves (FDCs) are an important tool for watershed management and their prediction in ungauged catchments is a challenging problem. Selecting the most appropriate model for prediction the FDC is itself a challenge that determines how theoretical improvements in prediction are transferred into engineering practice. Available performance metrics (e.g., Nash Sutcliffe Coefficient, error on flow moments) typically consider the aggregated ability of the model to predict all streamflow quantiles. These metrics may be inappropriate for model selection in practice because watershed management decisions are typically driven by a limited number of streamflow quantiles that may be poorly represented by an aggregate performance metric.

As an illustrative case study, the performance of three distinct FDC prediction approaches -- graphical, statistical and process-based -- are compared for ungauged streams in Nepal. The practical application of these predictions is to inform the design of run-of-river hydropower plants. The process-based approach provides the best prediction of the observed flow distribution and results in significantly higher Nash coefficients. However, the graphical approach provides the best prediction of the flow quantiles that are most relevant for hydropower design and reduces the design error caused by streamflow estimation. To assist in an application driven model selection process, we propose a novel model selection framework.