Deep and machine learning for daily streamflow estimation: a focus on LSTM, RFR and XGBoost

Abstract

Abstract Estimation accuracy of streamflow values is of great importance in terms of long-term planning of water resources and taking measures against disasters such as drought and flood. The flow formed in a river basin has a complex physical structure that changes depending on the characteristics of the basin (such as topography and vegetation), meteorological factors (such as precipitation, evaporation and infiltration) and human activities. In recent years, deep and machine learning techniques have attracted attention thanks to their powerful learning capabilities and accurate and reliable modeling of these complex and nonlinear processes. In this paper, long short-term memory (LSTM), random forest regression (RFR) and extreme gradient boosting (XGBoost) approaches were applied to estimate daily streamflow values of Göksu River, Turkey. Hyperparameter optimization was realized for deep and machine learning algorithms. The daily flow values between the years 1990–2010 were used and various input parameters were tried in the modeling. Examining the performance (R2, RMSE and MAE) of the models, the XGBoost model having five input parameters provided more appropriate results than other models. The R2 value of the XGBoost model was obtained as 0.871 for the testing set. Also, it is shown that deep and machine learning algorithms are used successfully for streamflow estimation.