Activity Number:
|
450
- Uncertainty Quantification for Environmental Applications
|
Type:
|
Topic Contributed
|
Date/Time:
|
Thursday, August 6, 2020 : 10:00 AM to 11:50 PM
|
Sponsor:
|
Section on Statistics and the Environment
|
Abstract #312643
|
|
Title:
|
Calibrating a WRF-Hydro Model Using a New Deep Learning-Based Calibration Method
|
Author(s):
|
Won Chang* and Saumya Bhatnagar and Jiali Wang and Seonjin Kim
|
Companies:
|
University of Cincinnati and Univ of Cincinnati - Cincinnati, OH and Argonne National Laboratory and Miami University
|
Keywords:
|
computer model calibration;
deep learning;
long short-term memory network;
data-model discrepancy;
hydrology;
WRF-hydro
|
Abstract:
|
The WRF-Hydro model enables high-fidelity simulation of streamflow in river basins by coupling high-resolution weather and hydrologic simulators. However, many input parameters for the model are highly uncertain, which poses a major obstacle in providing realistic flood forecasts based on the model. While the standard computer model calibration in principle can be applied to reduce the input parameter uncertainty, it suffers from identifiability issues for the effects from input parameters and data-model discrepancy and hence cannot successfully recover the true input parameter values even under mild model structure errors. To overcome this challenge we apply a new computer model calibration method based on a deep neural network (DNN) model equipped with long short-term memory (LSTM) layers for feature extraction. In a simulation study our approach was able to filter out the effects of data-model discrepancy and recover the true input parameter values thanks to the feature extraction capacity of the employed DNN model.
|
Authors who are presenting talks have a * after their name.