| Hydrologic time series forecasting is of great importance for water resources systems planning and operation. In the past decade, there have been significant advances in the development and applications of neural networks in various areas, such as medical science, computer science, and engineering. However, few applications of neural networks in hydrologic time series analysis are available. The purpose of this research is to test the applicability of neural networks in streamflow forecasting and data generation.;Three-layer (input, hidden, and output layers) feed-forward back-propagation neural networks were utilized in this study for forecasting river streamflow discharge. The following cases were analyzed: single-input single-output (SISO) series, multiple-input single-output (MISO) series, and multiple-input multiple-output (MIMO) series. Flow forecasting based on neural networks is compared with other forecasting procedures based on linear regression, canonical correlation, ARMA models, and transfer function models. The results demonstrated that neural networks can be quite useful for flow forcasting depending on the particular case. For instance, in case of SISO, MISO and MIMO monthly streamflow forecasting, neural networks improved the forecasting accuracy. The improvement can be explained by the non-linear nature of the analyzed relations, such as streamflow vs. combined snow-pack and temperature. In cases where the nature of the analyzed relation is linear or close to linear, such as current monthly streamflow versus previous monthly streamflow, neural networks performed equally well as the other models.;In the limited study on the application of neural networks in data generation, it is shown that neural networks have potential to be successfully applied to hydrologic data generation.;Redundancy in model parameters of neural networks is a controversial issue. While the traditional doctrine opposes the redundancy, some neural networks theoreticians consider the redundancy of neural networks an advantage because of their robustness. Through several applications, this research attempts to add some insight to the ongoing problem. |
