A machine learning algorithm for predicting naturalized flow duration curves at human influenced sites and multiple catchment scales

Abstract

Regional flow duration curves (FDCs) often reflect streamflow influenced by human activities. We propose a new machine learning algorithm to predict naturalized FDCs at human influenced sites and multiple catchment scales. Separate Meta models are developed to predict probable flow at discrete exceedance probabilities across catchments spanning multiple stream orders. Discrete exceedance flows reflect the stacking of k-fold cross-validated predictions from trained base ensemble machine learning models with and without hyperparameter tuning. The quality of individual base models reflects random stratified shuffling of spilt catchment records for training and testing. A Meta model is formed by retraining minimum variance base models that are bias corrected and used to predict final flows at selected percentiles that quantify uncertainty. Separate Meta models are developed and used to predict naturalised stochastic flows at other discrete exceedance probabilities along the duration curve. Efficacy of the new method is demonstrated for predicting naturalized stochastic FDCs at human influenced gauged catchments and ungauged stream reaches of unknown influences across Otago New Zealand. Important findings are twofold. First, independent observations of naturalised Median flows compare within few percent of the 50th percentile predictions from the FDC models. Second, the naturalised Meta models predict FDCs that outperform the calibrated SWAT model FDCs at gauge sites in the Taieri Freshwater Management Unit: Taieri at Tiroiti, Taieri at Sutton Creek, and Taieri River at Outram. Departures in the naturalised reference state are interpreted as flow regime changes across the duration curves. We believe these Meta models will be useful in predicting naturalised catchment FDCs across other New Zealand regions using physical catchment features available from the national data base.