Comparing a long short-term memory (LSTM) neural network with a physically-based hydrological model for streamflow forecasting over a Canadian catchment

Streamflow forecasting is crucial in water planning and management. Physically-based hydrological models have been used for a long time in these fields, but improving forecast quality is still an active area of research. Recently, some artificial neural networks have been found to be effective in simulating and predicting short-term streamflow. In this study, we examine the reliability of Long Short-Term Memory (LSTM) deep learning model in predicting streamflow for lead times of up to ten days over a Canadian catchment. The performance of the LSTM model is compared to that of a process-based distributed hydrological model, with both models using the same weather ensemble forecasts. Furthermore, the LSTM's ability to integrate observed streamflow on the forecast issue date is compared to the data assimilation process required for the hydrological model to reduce initial state biases. Results indicate that the LSTM model forecasted streamflows are more reliable and accurate for lead-times up to 7 and 9 days, respectively. Additionally, it is shown that the LSTM model using recent observed flows as a predictor can forecast flows with smaller errors in the first forecasting days without requiring an explicit data assimilation step, with the LSTM model generating a median value of mean absolute error (MAE) for the first day of lead-time across all forecast issue dates of 25 m3/s compared to 115 m3/s for the assimilated hydrological model.