Application of Levenberg-Marquardt Optimization Algorithm Based Multilayer Neural Networks for Hydrological Time Series Modeling

Abstract

It is very important to make hydrological time series modeling on water resource engineering and the decision making strategies of water resource management. In this paper, a comprehensive study on the application of Levenberg-Marquardt optimization algorithms based Multilayer Neural Networks in the monthly inflows of Demirkopru Dam, which is located in the Gediz Basin/Turkey, is presented. The best network structure which requires monthly areal precipitation, temperature and one month ahead areal precipitation values as the input data, is trained by using the 30 years monthly time series. The network was trained by the observed data set between 1977 and 1996, then tested by the data set observed between 1997 and 2006. The training and testing criteria which are selected as coefficient of determination (R2) and root-mean square error (RMSE) were calculated in the training period as 79.17 (%) and 31.95 (106 m3), respectively. These values are calculated in the testing period as 75.34 (%) and 34.37 (106 m3), respectively. Long term and seasonal term statistics show that the developed network simulates the inflows of Demirköprü Dam satisfactorily and the results represent that model can be used to estimate the monthly inflows. This was also proved with this study that Levenberg-Marquardt algorithm is the one of the fastest multilayer neural network training approaches and also a black box technique which is capable of reservoir inflow modeling without the detailing of the physical process.