Year: 2003
Type: conference-paper
Abstract: Many studies have used artificial neural networks (ANNs) for the prediction and forecasting of hydrological variables, including runoff, precipitation and river level, which are subsequently used for design or management purposes.However, although it is widely recognised that hydrological models are subject to parameter uncertainty, ANNs in this field have been almost exclusively deterministic with little attention paid to the uncertainty in the network weights.The inherent variability of hydrological processes means that no finite set of observations will give exact parameter values and therefore it is important to express network weights as a range of plausible values such that one, possibly incorrect, weight vector does not completely dominate the predictions.In this paper a synthetically generated data set is used as a tool for demonstrating the potential advantages of explicitly accounting for parameter uncertainty.A Markov chain Monte Carlo approach is used to sample from the distribution of possible network weights in an attempt to eliminate or reduce the potential problems that can be encountered during network training.By expressing the network weights as a distribution it is also possible to express the level of confidence with which ANN predictions are made.
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3093401227