Rainfall–Runoff Prediction at Multiple Timescales with a Single Long Short-Term Memory Network

Abstract

Abstract. Long Short-Term Memory Networks (LSTMs) have been applied to daily discharge prediction with remarkable success. Many practical scenarios, however, require predictions at more granular timescales. For instance, accurate prediction of short but extreme flood peaks can make a life-saving difference, yet such peaks may escape the coarse temporal resolution of daily predictions. Naively training an LSTM on hourly data, however, entails very long input sequences that make learning hard and computationally expensive. In this study, we propose two Multi-Timescale LSTM (MTS-LSTM) architectures that jointly predict multiple timescales within one model, as they process long-past inputs at a single temporal resolution and branch out into each individual timescale for more recent input steps. We test these models on 516 basins across the continental United States and benchmark against the US National Water Model. Compared to naive prediction with a distinct LSTM per timescale, the multi-timescale architectures are computationally more efficient with no loss in accuracy. Beyond prediction quality, the multi-timescale LSTM can process different input variables at different timescales, which is especially relevant to operational applications where the lead time of meteorological forcings depends on their temporal resolution.

Deep learning rainfall-runoff predictions of extreme events

“…The hypothesis tested in this work was that data-driven streamflow models are likely to become unreliable in extreme or outof-sample events. This is an important hypothesis to test because it is a common concern among physical scientists and among users of model-based information products (e.g., Todini, 2007), however prior work (e.g., Kratzert et al, 2019b;Gauch et al, 2021) demonstrated that data-based rainfall-runoff models were more reliable than other types of physically-based models, even in extrapolation to ungauged basins (Kratzert et al, 2019a). Our results indicate that this hypothesis is incorrect -the data-driven models (both the pure ML model and the physics-informed ML model) were better than benchmark models at predicting peak flows in almost all conditions, including extreme events and including when extreme events were not included in the training data set.…”

“…We used the same set of performance metrics that were used in previous CAMELS studies (Kratzert et al, 2019b(Kratzert et al, , a, 2021Gauch et al, 2021;Klotz et al, 2021). A full list of these metrics is given in Table 1.…”

“…These metrics depend heavily on the observed flow characteristics during a particular test period and, because they are less stable, are somewhat less useful in terms of drawing general conclusions. We report them here primarily for continuity with previous studies (Kratzert et al, 2019b(Kratzert et al, , a, 2021Frame et al, 2020;Nearing et al, 2020a;Klotz et al, 2021;Gauch et al, 2021), and because one of the objectives of this paper (Section 2.2) is to expand on the high flow (FHV) analysis by benchmarking on annual peak flows.…”

Rainfall–runoff prediction at multiple timescales with a single Long Short-Term Memory network

“…All code to reproduce our models and analyses is available at https://doi.org/10.5281/zenodo.4687991 (Gauch, 2021). The trained models and their predictions are available at https://doi.org/10.5281/zenodo.4071885 (Gauch et al, 2020a). Hourly NLDAS forcings and observed streamflow are available at https://doi.org/10.5281/zenodo.4072700 (Gauch et al, 2020b).…”

“…The trained models and their predictions are available at https://doi.org/10.5281/zenodo.4071885 (Gauch et al, 2020a). Hourly NLDAS forcings and observed streamflow are available at https://doi.org/10.5281/zenodo.4072700 (Gauch et al, 2020b). The CAMELS static attributes are accessible at https://doi.org/10.5065/D6G73C3Q (Addor et al, 2017b).…”