Usage of Artificial Neural Network for Generating Monthly Average Flows at the Ungaged River Profiles

This paper presents the results of a study that simulates historic average discharges by means of an Artificial Neural Network, within a variety of input data combinations. A cluster of small area watersheds in Eastern Serbia was selected for the study. The input data considered were: climatic data (precipitation, average air temperatures, vapor pressure and air humidity), physiographic, hydrogeological, soil, and vegetation indicators. A sequentially adaptive radial basis function network was applied for generating flow series. Sequential adaptation of parameters and structure were achieved using the extended Kalman filter. A criterion for network growth has been obtained from the Kalman filter’s consistency test. The best input data set and simulation results are presented and discussed for each watershed.