Hybrid light gradient boosting model-metaheuristic optimization framework with NASA POWER data for tropical streamflow prediction

Accurate streamflow prediction is vital for water resource management, particularly in data-scarce tropical regions. This study develops a hybrid streamflow prediction framework that integrates the Light Gradient Boosting Model (LGBM) with metaheuristic optimization, feature selection techniques, and NASA POWER meteorological data for the Muda River Basin, Malaysia. The LGBM performed best with the Grey Wolf Optimizer (GWO) and Sequential Forward Selection (SFS) at Syed Omar Bridge (NSE = 0.83, RMSE = 20.38 m3/s, WI = 0.95, PBIAS = +2.93%), whereas at Ladang Victoria it achieved the highest accuracy using the Teaching–Learning-Based Optimizer (TLO) and Boruta feature selection method (NSE = 0.84, RMSE = 33.54 m3/s, WI = 0.95, PBIAS = −1.00%). Compared with baseline models, the optimized LGBM frameworks consistently reduced prediction errors, improved efficiency metrics, and minimized systematic bias. These findings demonstrate the potential of open-source, data-driven, optimized machine learning models for reliable streamflow prediction in tropical basins.