Streamflow prediction combining WRF-Hydro modelling with machine learning

WRF-Hydro is one of widely used physics-based hydrological modeling framework integrating various terrestrial and hydrological processes to simulate water, energy and momentum fluxes from land to the atmosphere as well as streamflow. The parameters of WRF-Hydro needs to be calibrated as they are employed to simplify hydrological and physical processes for different regions and scales. However, even calibrated regional parameters are not sufficient to overcome a gap between model outputs and observations due to uncertainties caused by generalized and low-dimensional inputs in the model. In this study, we therefore developed the framework, combining Long Short-Term (LSTM) networks, one of machine learning approaches, with WR F-Hydro modelling to improve WRF-Hydro output by reducing uncertainty-induced errors. LSTM is an artificial recurrent neural network architecture which has enhanced long-term memory enables higher accuracy prediction with multiple time-series data. With the LSTM structure, machine was trained to capture feature of underlying discrepancy between the model outputs and obervations. With the trained LSTM, the future errors between the model outputs and observations are predicted, and then applied to improve the model predictions. For case study, we used inflow observation in Soyang river dam basin, South Korea for year of 2013 to 2018 to evaluate the performance of our framework. First of all, the Parameter EStimation Tool (PEST) is employed to calibrate the WRF-Hydro in the study area. With this calibrated model, we generated 6 year long inflow which were used for machine learning (5 years for training set and last year for validation set). Based on our evaluation with the validation set, we show that the combining WRF-Hydro modelling with LSTM is able to improve the streamflow prediction from prediction with WRF-Hydro only.

Keyword : WRF-Hydro, Machine learning, LSTM, PEST

This work was supported by Korea Environmet Industry & Technology Institute(KEITI) through Advanced Water Management Research Program, funded by Korea Ministry of Environment(83089) and National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (2018R1A1A3A04079419)