The Role of Metaheuristic Algorithms in Tuning Extreme Learning Machine Model for Lake Water Level Modeling

Abstract Lake water level (WL) is an important indicator of the lake’s physical condition, and its variation may significantly influence ecosystem stability and water supply security. Hence, precise estimation of lake WL is important for environmental protection and sustainable water resource management. Consequently, the current study introduced a novel hybridized extreme learning machine (ELM) with four optimization algorithms, namely grey wolf optimizer (GWO-ELM), marine predator algorithm (MPA-ELM), mountain gazelle optimizer (MGO-ELM), and whale optimization algorithm (WOA-ELM). The performance of the developed hybrid models was compared with the standalone ELM and multilinear regression (MLR) using different statistical metrics such as coefficient of determination (R 2 ), Root mean square error (RMSE), Nash Sutcliffe Efficiency (NSE), and mean absolute error (MAE). The results demonstrated that the hybrid models outperformed the standalone models. During the testing phase, the best prediction results were obtained using the MGO-ELM model at Lake Hawassa (R 2 = 0.997, RMSE = 0.022 m, MAE = 0.008 m, and NSE = 0.996), Langano (R 2 = 0.987, RMSE = 0.056 m, MAE = 0.022 m, and NSE = 0.986), Abiyata ((R 2 = 0.999, RMSE = 0.02, MAE = 0.012 m, and NSE = 0.998) and Ziway (R 2 = 0.995, RMSE = 0.036 m, MAE = 0.021 m, and NSE = 0.994). The MGO-ELM model enhanced the estimation performance in comparison with ELM and MLR models by 70.322%-89.848% and 44.578%-82.609%, respectively, in terms of testing set RMSE. Generally, the developed hybrid ELM models can be used as a practical computational method for sustainable water resource management in the Ethiopian Rift Valley Lakes.