LSTM-Based Rainfall–Runoff Modeling at Arbitrary Time Scales

In recent years, rainfall–runoff models based on machine learning techniques, particularly Long Short-Term Memory (LSTM) networks, have proven highly successful. They outperform conceptual hydrologic models, predict multiple basins in a single model, and allow for predictions in ungauged basins (Kratzert et al. 2018, Kratzert et al. 2019, Kratzert et al. 2019a). Yet, there remain open challenges toward operational use of such models. One major challenge is the fact that most research so far has focused on machine learning for daily predictions. While daily predictions are highly relevant for medium- to long-range forecasts, they are too coarse to capture characteristics such as the precise timing of peaks in short-range forecasts. Hence, streamflow predictions at sub-daily time scales are a key ingredient for operationally usable machine learning models. To this end, we demonstrate a novel approach that can generate predictions at arbitrary temporal frequencies in a single LSTM-based model. For instance, a single model can generate hourly, three-hourly, and daily predictions, each up to a different temporal horizon. Moreover, the model can ingest different forcing products (or other input variables) for each time scale, which is important since high-frequency forcings usually have a shorter forecast horizon than lower-frequency forcing products. To test our proposed model, we train a single LSTM-based model on NLDAS forcings and USGS streamflow data from 516 basins across the contiguous United States, aggregated to time scales between one hour and one day. Preliminary results indicate that this technique outperforms state-of-the-art conceptual hydrologic models and its accuracy does not degrade compared to a daily LSTM, which can only predict daily streamflow.