Daily streamflow modeling using hybrid models based on random forest and gradient boosting integrated with Harris Hawks and Jaya optimization algorithms

Accurate prediction of streamflow using reliable and cost-effective techniques is paramount for watershed management. The current study aimed to predict streamflow of Katar River, Ethiopia using different standalone and hybrid models. The modeling was carried out using Random Forest (RF), gradient boost regression (GBR) and least square support vector machine (LSSVM) models. Subsequently, four novel hybrid models were developed by conjugating RF and GBR with Harris Hawks optimization (HHO) and Jaya Optimization Algorithms (i.e., RF-Jaya, RF-HHO, GBR-Jaya, and GBR-HHO). Ten years of daily streamflow and rainfall data were utilized for calibrating and validating the models. Several statistical metrics such as determination coefficient (R²), Nash–Sutcliffe efficiency (NSE), RMSE-observation standard deviation ratio (RSR), Root mean square error (RMSE), Percent bias (Pbias), and visually-based evaluation metrics were utilized to compare the efficiency of the models. The result revealed that RF surpassed LSSVM and GBR with the validation period NSE = 0.915, R² = 0.916, RMSE = 5.943 m³/s, Pbias = −0.939% and RSR = 0.085. The result of the study also demonstrated that, in the validation period, the best streamflow prediction was obtained from hybrid RF-Jaya (RMSE = 4.976 m³/s, NSE = 0.94, R² = 0.94, RSR = 0.059 and Pbias = 0.624%). Generally, the Jaya and HHO optimization methods demonstrated improved streamflow results compared to the standalone models in the study catchment.