Application of Meta-Models for Accurate Calibration of Hydrological Model Parameters

Abstract

One of the impacts of climate change is an increase in the frequency of floods. The efficient
and optimized flood analysis system needs to be used for the reliable flood forecasting. The
credibility and the reliability of the flood forecasting system is depending upon the framework
used for its parameter optimization. A comprehensive framework for optimizing the input
parameters of the computationally extensive distributed hydrological model has been presented.
A large river basin has the high spatio-temporal heterogeneity of aquifer and surface properties.
Estimating the parameters in fully distributed hydrological model is a challenging task. The
parameter optimization becomes computationally more demanding when the model input
parameters (30 to 100 even greater) have multi-dimensional parameter space, many output
parameters which make the optimization problem multi-objective and large number of model
simulations requirement for the optimization. Aforementioned challenges are met by introducing
the methodology to optimize the input parameters of fully distributed hydrological model,
following steps are included (1) screening of the parameters through Morris sensitivity analysis
method in different flow periods, so that optimization would be performed for sensitive
parameters, different scalar output functions are used in this regard (2) Surrogate models or
meta-models are used to simulate the hydrological response of a dynamic model (3) sampling of
parameters values using the optimized ranges obtained from the meta-models developed from
multivariate regression adaptive splines (MRAS); the results are evident that the parameter
optimization using the proposed framework is efficient can be effectively performed. The
efficiency and performance of the proposed framework has been demonstrated through the
accurate calibration of the model with fewer model runs. This study also demonstrates the
importance and use of scalar functions in calculating sensitivity indices, when the model output
is temporally variable.