Sensitivity analysis of constituent generation parameters of an integrated hydrological and water quality model using a GMDH polynomial neural network
MODSIM · 2017
Catchment water quality models are notoriously over-parametrised.Given this condition, it is useful to be able to identify which parameters have the greatest influence on the model results.In theory, this could be accomplished through a detailed first principles interrogation of the mathematical structure of the model in an abstract manner.This however, is impractical in most instances owing to the complexity of the models and posteriori methods of parameter sensitivity analysis are more conventional.As with most aspects of large-scale modelling endeavours, a major consideration in choosing a technique for sensitivity analysis is efficiency and a compromise between computational effort and numerical accuracy is usually negotiated.ANOVA based sensitivity analysis methods are very popular as they offer a holistic survey of the parameter sensitivity by not only accounting for the response of the model output surface due to the activity of single parameters acting independently, but also due to the interaction between parameters.These global sensitivity indices are usually calculated by Monte Carlo simulation and may be too computationally demanding to be routinely applied in water quality modelling scenarios.We demonstrate the application of the group method of data handling (GMDH) inductive, self-organising modelling method to the sensitivity analysis of constituent generation parameters of an integrated hydrological and water quality model.By using a modestly sized sample input-output dataset, a GMDH neural network is used to synthesise a sparse, random-sampling high dimensional model representation (RS-HDMR) that can be used to calculate first and second order Sobol sensitivity indices.This algorithm potentially leads to reductions in computational cost of 2-3 orders of magnitude over Monto Carlo simulation.Although several other adaptive methods for efficiently constructing a sparse RS-HDMR have been reported in the literature, such as polynomial chaos expansions, the parameter selection and noise filtering characteristics of the GMDH network may result in more optimal HDMR expansion.