Learning time reduction in artificial neural networks applied to hydrology

Abstract

One problem of artificial neural networks (ANN) used to forecast and model rainfall and runoff phenomena is the time required to obtain a good learning. This is due to slow algorithms and to the large number of parameters in networks. This paper focuses on reducing the learning time of ANN by selecting an adequate algorithm and by reducing the number of parameters. Specifically, Principal Component Analysis (PCA) and 4 families of algorithms were examined. PCA selects the best data projection of input. information, in order to represent a significant proportion of sample variance with a minimum of independent variables or principal components. This produces fewer and independent variables and reduced learning times due to a smaller number of model parameters. The Adaptive gradient-descent momentum, Conjugate gradient-descent, Broyden-Fletcher-Goldfard-Shanno algorithm and Levenberg-Marquardt algorithms were examined. The procedure chooses the algorithm which requires smaller learning times to obtain the best fit of the hydrologic information. Two hydrologic problems were used to test the algorithms and the PCA technique. The first one forecasts spring and summer flows of the San Juan river in Argentina, and the second one predicts rainfall intensities in northern California. Results show that PCA produces faster learning periods without diminishing fitness quality and that the Levenberg-Marquardt algorithm is the best choice for feed-forward networks. The authors recommend the use of the Levenberg-Marquardt algorithm for training feed-forward artificial neural networks and the use of principal component analysis in models with a large number of correlated explanatory variables.