Generalized regression neural network in monthly flow forecasting

Abstract

The majority of the artificial neural network (ANN) applications to water resources data involve the employment of the feed forward back propagation method (FFBP). In this study, an ANN algorithm, generalized regression neural network (GRNN), was employed in monthly mean flow forecasting. The performances of the GRNN and the FFBP methods were compared initially for forecasting of monthly mean river flows and training the neural networks using the observed data; then the forecasting study was carried out using the AR model-generated synthetic monthly mean flow series for training stage. The GRNN simulations did not face the frequently encountered local minima problem of the FFBP applications and did not generate forecasts that are physically implausible. It was seen that FFBP forecasting performance was sensitive to the randomly assigned initial weights. This problem, however, did not occur in the GRNN simulations. The GRNN approach does not require an iterative training procedure, unlike the FFBP method. GRNN forecasting performance was found to be superior to the FFBP, statistical, and stochastic methods in terms of the selected performance criteria.