Exploring the added value of machine learning methods in predicting flow duration curves a comparative analysis for ungauged catchments

2016-12-12
Doğulu, Nilay
Kentel Erdoğan, Elçin
Understanding catchment hydrology is a fundamental concern for hydrologists and water resources planners. In this context, given the increasing demand for streamflow information at sparsely gauged or ungauged catchments, there has been great interest in estimating flow duration curve (FDC) due to its many practical applications. Statistical methods have been widely used for the modelling of FDCs at ungauged sites. These methods usually rely on estimation of flow quantiles, or quantitative characteristics of the FDCs representing their shape such as slope and parameters of statistical distribution, often in the context of regionalization. However, there are limited studies using methods of machine learning. Potential of various machine learning approaches for estimating FDCs is yet to be explored although these methods have successfully and extensively applied to solve various other water resources management and hydrological problems. This study addresses this gap by presenting a comparative performance evaluation of the methods: i) Multiple Linear Regression (MLR), ii) Regression Tree (RT), iii) Artificial Neural Network (ANN), iv) Adaptive Neuro-Fuzzy Inference System (ANFIS). Comparison of these methods is done for FDCs of the Western Black Sea catchment in Turkey modelled by relating flow quantiles to a number of variables representing catchment and climate characteristics. Accuracy of predicted FDCs is assessed by three different measures: the Root Mean Squared Error (RMSE), the Nash-Sutcliffe Efficiency (NSE) and the Percent Bias (PBIAS).
Citation Formats
N. Doğulu and E. Kentel Erdoğan, “Exploring the added value of machine learning methods in predicting flow duration curves a comparative analysis for ungauged catchments,” presented at the American Geophysical Union Fall Meeting, 2016, San Francisco, USA, 2016, Accessed: 00, 2021. [Online]. Available: https://hdl.handle.net/11511/76494.