On the Development of Improved Artificial Neural Network Model and Its Application on Hydrological Forecasting

As conventional multilayer backward-propagation network does not perform well on parameter estimation and convergence, several improved backward-propagation algorithms, such as VLBP, MOBP, CGBP and LMBP, were developed. In order to investigate simulation performance of each algorithm to construct the BP network model suitable for hydrological forecasting, five backward-propagation (BP) neural networks which are based on different algorithms are trained and compared among them. The results of experiments show that the Levenberg-Marquardt backpropagation (LMBP) neural network with a Levenberg-Marquardt based algorithm with enhanced optimization performance has better system identification capacity and is suitable for network in which performance index is evaluated with mean- square error. Therefore, LMBP neural network are chosen for construction of hydrological forecasting model. The flood forecast results compare well with observed data. According to criterion, the model can be used as a favorable method and can be applied in other nonlinear system identifications.