Two Stage History Matching for Hydrology Models via Machine Learning

EasyChair Preprint 1040

Abstract

The reliability of a hydrological model (HydM)  highly depends on how well the model parameters are estimated through the history matching (HisM) process. Direct HisM (DHisM) that calibrate input parameters by iteratively executing a model is widely applied in water resource estimation. The computational time of DHisM is prohibitive, as a single run on the model may take several hours. In practice calibration accuracy is compromised to arrive at a solution. Therefore, it is desirable to develop a proxy model that can replace hydrological model in the HisM process. In this study, we propose a two stage HisM, wherein  we first  develop a proxy model for HydM using artificial neural network  techniques. Next we apply, ant colony optimisation (ACO_R) and robust parameter estimation (ROPE) methods for calibrating the parameter of HydM.  This methodology is illustrated for the Dandalups catchment of Western Australia to  calibrate five global parameters of Land Use Change Incorporated Catchment (LUCICAT) by matching  33 annual daily streamflow peaks. The results reveal that the replacing the LUCICAT by proxy model reduces the  computational time by more than 90% with similar accuracy to DHisM and show higher consistency (via standard deviation of RMSE) and reduction of parameter uncertainty compared to DHisM.

Keyphrases: History matching methods, Proxy Model, history matching, machine learning

@booklet{EasyChair:1040,
  author    = {Dewi Tjia and Ritu Gupta and Muhammad Alam},
  title     = {Two Stage History Matching for Hydrology Models via Machine Learning},
  howpublished = {EasyChair Preprint 1040},
  year      = {EasyChair, 2019}}