Estimation performance of neural networks

In order to test neural network abilities as estimators of engineering value, a network is presented to derive streamflow from precipitation data. Validation tests show good performance, hence increasing confidence in these methods. Monthly mean squared errors remaining after adjustment are presented and compared with those of deterministic methods, since these are other options for the estimation problem. A possible caveat of artificial neural networks (ANN) is that they are very difficult to interpret. Interpretation of the learnt representation in this case is offered by simulating with selected inputs, showing reasonable results and providing some insight in the hydrologic process being modeled. This is a generic possibility for dealing with black-box models. When estimating some system´s behavior it is interesting to know whether the qualitative representation is also faithful. In the proposed example, special properties of the flow series with significance in hydrology, such as ranges, are obtained and compared with the sample values, along with other statistical features.