Uncertainty Analysis on Neural Network Based Hydrological Models Using Probabilistic Point Estimate Method

Modeling hydrological processes are always a challenge due to incomplete understanding of the physics of the process. Therefore, various levels of simplification are essential during modeling, which are otherwise very complex. In addition, most of the hydrological processes being natural are random processes. Apart from the standard physics based models developed in hydrology, the artificial neural network (ANN) approach has been getting lot of attention plausibly due to the complexity associated with the system. However, in most of the application of ANN in hydrology the model is considered as deterministic despite a large amount uncertainty associated with the final ANN models. Very recently, there has been considerable interest to quantify the uncertainty associated with ANN models, and not much work is reported since application of standard methods for uncertainty quantification was difficult due to the parallel computing architecture of the ANN. This paper presents the application of probabilistic point estimate in quantifying the uncertainty of ANN river flow forecasting model. The method is demonstrated through a case study of L’Anguille watershed located in United States. The results show that the method effectively quantifies uncertainty in the model output by estimating the parameters in orthogonal domain. The study also suggests that the method can be employed for models with lesser number of simulations, and do not require much knowledge about the parametric distribution of the model.