Hydra-LSTM: A Semi-Shared Machine Learning Architecture for Prediction across Watersheds

Abstract Long short-term memory (LSTM) networks are used to build single models that predict river discharge across many catchments. These models offer greater accuracy than models trained on each catchment independently, if the same variables are used as inputs for each catchment. However, the same data are rarely available for all catchments. This prevents the use of variables available only in some catchments, such as historic river discharge or upstream discharge. The only existing method that allows for optional variables requires all variables to be in the initial training of the model, limiting its transferability to new catchments. To address this limitation, we develop the Hydra -LSTM. The Hydra -LSTM is able to use some variables across all catchments to make predictions and use further variables in other catchments where they are helpful and available. This allows general training and the use of catchment-specific data. The bulk of the model can be shared across catchments, maintaining the benefits of multicatchment models to generalize while also benefiting from using bespoke data. We apply this methodology to 2-day-ahead river discharge prediction in the western United States, a small enough time step to expect our models to be skillful and difficult enough to expect differences between models. We obtain more accurate quantile predictions than multicatchment and single-catchment LSTMs while allowing forecasters to introduce and remove variables from their prediction set. We test the ability of the Hydra -LSTM to incorporate catchment-specific data, introducing historical river discharge as a catchment-specific input, outperforming other commonly used models.