A Simplified Approach for Rainfall-runoff Modeling Using Advanced Soft-computing Methods

This study investigates three modeling approaches based on Xtreme Boosting Machine (XGBoost), Genetic Algorithm-optimized Emotional Neural Network (GA-EmNN) and Group Method of Data Handling-Neural Network (GMDH-NN) for rainfall-runoff modeling. The redundancy capability of Principal Component Analysis (PCA) was applied to solve the problems of input selection in these machine learning models. Hence, the objective of this study is to develop an efficient approach to pre-process the data structure and ensemble it with machine learning models that produce a higher predictive accuracy. The three XGBoost, GA-EmNN and GMDH-NN models are trained and validated for streamflow forecasting using monthly rainfall and monthly runoff data of the Jhelum basin, India. Statistical fitness indices, like correlation (r), Root Mean Square Error (RMSE), relative Nash-Sutcliffe Efficiency coefficient (r NSE), Percent Bias (PBias) and Kling-Gupta Efficiency (KGE), were used for model performance assessment. The analysis of results revealed that GA-EmNN model was well capable (Training: R=0.999, RMSE = 345.48 cumec, Testing: R=0.997, RMSE = 428.35 cumec) of predicting the streamflow, followed by GMDH-NN and XGBoost. The results of this approach would benefit future modeling efforts, by underlining the application of these methods in hydrological modeling.