Are long short-term memory (LSTM) model simulations of watershed discharge improved when water storage is included as input? Case Study in Rum River Watershed, MN

Flooding is one of the most financially devastating natural hazards in the world. Accurate forecasting of floods is important to reduce the financial costs and loss of human lives. We hypothesize that utilizing the relation between watershed storage and watershed discharge has the potential to improve existing flood forecasting systems.  This presentation assesses the relation between daily water storage (S) and discharge (Q) with physically-based hydrological modeling, and storage-discharge dynamics simulated by the machine-learning (ML) algorithms for the Rum River Watershed, a HUC-8 watershed located in East-Central Minnesota. Used for predicting the outputs that represent arbitrary non-linear functions between predictors and predictands, ML can improve the representation  of the non-linear relation between water storage and discharge. In the present study the Long Short-Term Memory (LSTM), the time-series deep learning neural network used for predicting rainfall-runoff relations,  is used for simulating the non-linear relation between S and Q. The temporally- and spatially-distributed discharge and water storage was simulated for the 131 subwatersheds (HUC-12) composing the Rum River Watershed with a calibrated Hydrological Simulated Program-Fortran (HSPF) model. Water storage output was discriminated into the components of soil moisture, deep groundwater, active groundwater, interflow, interception, and surface runoff. The meteorological data used in the simulation were applied to the LSTM model with the simulated discharges as the output, and the simulated distributed storage as a state variable. For these simulations 66% of the inputs were used for training the LSTM model and 33% were used for testing the training model. The result of the LSTM simulation shows that by including the total water storage, and any of the components of total water storage, the prediction of discharge at the mouth of the Rum River Watershed is improved. Furthermore, by downscaling the information to the 131 subwatersheds, by including the water storage for each of the subwatersheds, the discharge at the mouth of each subwatershed is improved, and so is the discharge at the mouth of the Rum River watershed. 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 watersheds.