Machine learning method is an alternative for the hydrological model in an alpine catchment in the Tianshan region, Central Asia

Abstract

Study regionKaidu River catchment in the Tianshan Mountain, northwestern China. Study focusThis paper compared the applicability and accuracy of four machine learning models and two hydrological ones to simulate the daily streamflow and extreme streamflow of the Kaidu River catchment. The machine learning models are Support Vector Regression (SVR), eXtreme Gradient Boosting (XGBoost), Random Forests (RF), and Long Short-Term Memory (LSTM), while the hydrological models are the Soil and Water Assessment Tool (SWAT) and the extended SWAT with a glacier dynamic module (SWAT-Glacier). New hydrological insights for the regionLSTM achieved better model performance in simulating daily streamflow than SWAT and SWAT-Glacier, with Kling-Gupta efficiency of 0.92, 0.82, and 0.80, respectively. Meanwhile, SVR, XGBoost, and RF showed satisfactory performance, with KGE of 0.67, 0.71, and 0.70, respectively. LSTM, SWAT and SWAT-Glacier could well simulate the annual peak flow (i.e., annual maximum 1-day streamflow and 5-day average streamflow) but failed to mimic the annual minimum 7-day average streamflow, with PBIAS exceeding 28%. Furthermore, all the models failed to reproduce the dates of hydrological extremes. Nevertheless, using the quantile loss function in the LSTM model resulted in significantly improved model performance in the low streamflow indices, compared to that using mean squared error as the loss function. Overall, LSTM could be a good alternative for simulating daily streamflow and extreme streamflow in data-scarce catchments.