Assessment of machine learning models for short-term streamflow estimation: the case of Dez River in Iran

Accurate streamflow prediction is indispensable for efficient water resources management. In recent years, numerous investigations have utilized artificial intelligence (AI) and machine learning (ML) approaches for forecasting streamflows. The objective of this study is to assess eight AI techniques for predicting river flows. The ML models include adaptive neuro fuzzy inference system (ANFIS), support vector regression (SVR), M5P model tree, adaptive boosting (AdaBoost), genetic programming (GP), gradient boosting regression (GBR), extreme gradient boosting regression (XGBoost), and K-nearest neighbors (KNN). The daily discharges of the Dez River measured at the Telezang station between 2011 and 2022 were used in this study. Based on the obtained results, ANFIS outperformed other ML models examined in this study based on six criteria. Furthermore, the uncertainty analysis was conducted. The results demonstrated that the ANFIS model achieved the best river flow estimations, followed by the GBR model. In terms of the reliability of the experimental dataset, ANFIS and GBR indicated outstanding results, achieving uncertainty percentages of 96.77 and 93.55, respectively, signifying their excellent performance. It is postulated that the achieved results not only can be exploited as an input for hydrological modeling but also can help authorities to conduct better river management.