[A-HW20] Hydrology & Water Environment

[AHW20-06] Toward Downscaling Storage-Discharge Dynamics: Training Long Short-Term Memory (LSTM) model for Simulating Nonlinear Storage-Discharge Relation in HUC-8 Rum River Watershed, MN

*Pai-Feng Victor Teng¹ (1.University of Minnesota)

Keywords:Hydrological Modeling, Machine Learning , Long Short-Term Memory , Hydrology

Flooding is one of the most financially devastating natural hazards in the world. The current hydrological models have focused on rainfall-runoff models in flood forecasting, and studying storage-discharge relations has the potential to improve existing flood forecasting. This presentation will assess the relation between daily water storage (ΔS) and discharge (Q) with physical-based hydrological modeling, and storage-discharge dynamics with the machine-learning mechanism located at Rum River Watershed, a HUC8 watershed in Minnesota, between 1995-2015. Currently, linear regression models cannot adequately predict the relationship between the total ΔS and total Q at the HUC-8 watershed (R² = 0.3667). Since machine learning (ML) algorithms have already been used for predicting the outputs that represent arbitrary non-linear functions between predictors and predictands, this research will determine how ML algorithms will be used for improving the accuracy of the non-linear relation of the storage-discharge dynamics. Long Short-Term Memory (LSTM), the time-series deep learning neural network that has been used for predicting rainfall-runoff relations, will be used for simulating non-linear relations between ΔS and Q. It will compare two sets of storage-discharge relationships with the hydrological variables simulated by the semi-distributed Hydrological Simulated Program-Fortran (HSPF): dynamics between simulated discharge and input hydrological variables, including air temperature, cloud cover, dew point, potential evapotranspiration, precipitation, evapotranspiration, solar radiation, and wind speed, and the dynamics between simulated discharge and input variables that also includes total water storage at the HUC-8 watershed. 7670 samples are used. 90% of the inputs will be used for training the LSTM network and 10% will be used for testing the prediction. The result shows that the inclusion of total water storage can improve the prediction of total discharge at Rum River Watershed. Inputs of LSTM network can already adequately predict the discharge (NSE = 0.6147, R² = 0.8801, Bias Test = -0.4127) without including total water storage; inputting total water storage can significantly improve the prediction (NSE = 0.6977, R² = 0.9652, Bias Test = -0.4127). Yet, the main challenge is that both results underpredicted the discharge values in the spring of 2014. The result of the research lays the foundation for assessing the accuracy of downscaling storage-discharge dynamics by applying LSTM networks to evaluate storage-discharge dynamics at smaller, HUC-12 watershed simulated by hydrological models.