Comparison of novel approaches to uncertainty analysis of hydrological models
Comparison of novel approaches to uncertainty analysis of hydrological models
Mr. Nagendra Kayastha
Abstract
Hydrological models are simplified representations of hydrological behaviour of a catchment. The fact that hydrological processes are described by mathematical equations and the corresponding parameters are estimated using observations leads to uncertainties in model output. The uncertainty stems from the parameters, model structure and measurements of input and output data. The quantification of that uncertainty gives important information to the decision makers.
The conventional type of parameter uncertainty analysis methods in hydrological modelling is impracticable to run in real time due to high computational demand. It is necessary to explore the alternative solutions for this problem. In this thesis, we proposed a novel method for quantifying uncertainty in hydrological model using machine learning techniques named as MLUE. Uncertainty of the model outputs is estimated in terms of uncertainty bound which is resulted from Monte Carlo simulation (MCS). The uncertainty bound is expressed in the form of two quantiles representing prediction intervals (PIs).
Machine learning techniques such as linear regression, (LR), model trees (MT) and artificial neural networks (ANN) were used to approximate functional relationship between input data and uncertainty output and constructed model was validated by estimating the prediction intervals for unseen (test) data. In addition to the classical performance measure such as root mean squared error and coefficient of correlation, prediction intervals coverage probability (PICP) and mean prediction interval (MPI) are computed.
The applicability of the technique is studied by doing a case study in Brue catchment in UK using a lumped conceptual HBV-IHE 2007 model.
The results obtained for the case study suggested that the machine learning technique could provide useful uncertainty outputs that give engineers more information for design or decision making purposes.
It can be concluded that simpler and easier modelling approach like machine learning techniques can provide quick and cost effective solution for parameter uncertainty problems in hydrological models such cases where the main concern is the prediction of the magnitude of variable rather than explicitly dealing with the complex system.