 |
| Titel |
A novel approach to parameter uncertainty analysis of hydrological models: Application of machine learning techniques |
| VerfasserIn |
D. L. Shrestha, N. Kayastha, D. P. Solomatine |
| Konferenz |
EGU General Assembly 2009
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 11 (2009) |
| Datensatznummer |
250023132
|
|
|
|
|
|
| Zusammenfassung |
| Monte Carlo (MC) simulation-based techniques are widely used for analyzing parameter
uncertainty in hydrological models. Although MC simulations are flexible and robust,
and capable of solving a great variety of problems, they are not always practicable
for computationally intensive models. This study presents a novel approach for
assessment of parameter uncertainty in hydrological models using machine learning
techniques. The presented approach replicates MC simulation by using various machine
learning techniques, which is subsequently used for assessment of model parametric
uncertainty. It is assumed a hydrological model M(p) is given and the propagation of the
uncertainty in parameters p to the output is to be investigated. MC simulation of model
M(p) is run and the stored realizations are used to form the dataset for training
machine learning models. One of the issues was selection of the input variables for
the machine learning models; it was done by searching for the variables (or their
transformed variants) with the highest relatedness (average mutual information) to the
sought distribution of the model M output. Machine learning models are trained
to approximate the functional relationships between the variables characterizing
the process modelled by M(p) and the uncertainty descriptors of its output. The
trained machine learning models encapsulate the underlying characteristics of the
parameter uncertainty and can be used to predict uncertainty descriptors for the new
data.
In this study three machine learning models - artificial neural networks, model trees and
locally weighted regressions are used. The approach was demonstrated by estimating
parameter uncertainty of a lumped conceptual hydrological model, HBV with application to a
case study of meso scale mountainous catchment of Nepal. Uncertainty measures such as
prediction intervals estimated by three machine learning methods are compared to those
obtained by MC simulation in verification period. The results are promising as the uncertainty
measures estimated by machine learning models are reasonably accurate. The proposed
technique could be useful in real time applications for computationally intensive models
(e.g. physically based hydrological models) which require run times that make
traditional MC analysis impractical and when the forecast lead time is very short. |
|
|
|
|
|
|