Exploring the ability of LSTM-based hydrological models to simulate streamflow time series for flood frequency analysis

An increasing number of studies have shown the prowess of long short-term memory (LSTM) networks for hydrological modelling and forecasting. One drawback of these methods is the requirement for large amounts of training data to properly reproduce streamflow events. For maximum annual streamflow, this can be problematic since they are by definition less common than middle or low flows, leading to under-representation in the model’s training set. This study investigates six methods to improve the peak-streamflow simulation skill of LSTM models used for flood frequency analysis (FFA) in ungauged catchments. These include adding meteorological data variables, providing streamflow simulations from a distributed hydrological model, oversampling peak-streamflow events, adding multi-head attention mechanisms, adding data from a large set of “donor” catchments, and combining some of these elements in a single model. Furthermore, results are compared to those obtained by the distributed hydrological model HYDROTEL. The study is performed on 88 catchments in the province of Quebec using a leave-one-out cross-validation implementation, and an FFA is applied using observations, as well as model simulations. Results show that LSTM-based models are able to simulate peak streamflow as well as (for a simple LSTM model implementation) or better than (with hybrid LSTM–hydrological model implementations) the distributed hydrological model. Multiple pathways forward to further improve the LSTM-based model’s ability to predict peak streamflow are provided and discussed.