Comparison of Levenberg-Marquardt and Bayesian Regularization Learning Algorithms for Daily Runoff Forecasting

Abstract

In this study, Multilayer Perceptron (MLP) with Levenberg-Marquardt and Bayesian Regularization algorithms machine learning methods are compared for modeling of the rainfall-runoff process. For this purpose, daily flows were forecast using 5844 discharge data monitored between 1999 and 2015 of D21A001 Kırkgöze gauging station on the Karasu River operated by DSI. 6 scenarios were developed during the studies. Our findings indicate that the estimated capability of the Bayesian Regularization algorithm were close to with Levenberg-Marquardt algorithm for training and testing, respectively. This study shows that different network structures and data representing land features can improve prediction for longer lead times. We consider that the ANN model accurately depicted the Karasu flows, and that our study will serve as a guide for more research on flooding and water storage.