آب و فاضلاب (Mar 2009)
Effect of Input Variables Preprocessing in Artificial Neural Network on Monthly Flow Prediction by PCA and Wavelet Transformation
Abstract
River flow forecast has of long been the focus of attention due to its wide applications in water-related sciences. Development of new models and advanced techniques will bring about drastic changes in the estimation of this dynamic and nonlinear system. In this research, feed-forward Artificial Neural Network (ANN) was used to predict monthly flow. Given the numerous flow forecast variables used in the present study, identification of variables effective in the network was necessary to help obtain improved results. For this purpose, we modeled the flow using the Principal Component Analysis (PCA) technique that reduces the number of input variables to include only the ones effective in ANN (PCA-ANN). PCA was first employed to reduce the number of input variables whereby 18 original variables were changed to 18 new components and the first 8 in the best model were then selected as network inputs. In addition, wavelet transformation was used for preprocessing input variables in the network to develop a model for flow forecasting (WNN). Comparison of the results obtained from the three models (ANN, PCA-ANN, and WNN) indicated the positive effect of preprocessing by wavelet and PCA on input variables. Another finding of the study was that the proposed model (PCA-ANN) had a simpler network architecture, faster training speed, and more satisfactory predicting performance in comparison with ANN and WNN models.