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

Abstract

Abstract. Long Short-Term Memory (LSTM) networks have been applied to daily discharge prediction with remarkable success. Many practical applications, however, require predictions at more granular timescales. For instance, accurate prediction of short but extreme flood peaks can make a lifesaving 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 difficult 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 different temporal resolution than more recent inputs. In a benchmark on 516 basins across the continental United States, these models achieved significantly higher Nash–Sutcliffe efficiency (NSE) values than the US National Water Model. Compared to naive prediction with distinct LSTMs 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.

Benchmarking data-driven rainfall–runoff models in Great Britain: a comparison of long short-term memory (LSTM)-based models with four lumped conceptual models

“…The blue line shows the histograms for those stations between the two dashed lines. from as many catchments as possible (Gauch et al, 2021b). It is important to interpret the number of parameters for each model type in light of this fact.…”

“…These models range from physically based, spatially explicit models such as SHETRAN (Birkinshaw et al, 2010), CLAS-SIC (Crooks et al, 2014) and PARFLOW (Maxwell et al, 2009) to lumped conceptual models such as TOPMODEL (Beven and Kirkby, 1979) and VIC (Liang, 1994). Additionally, data-driven models have also been used for modelling rainfall-runoff processes (Reichstein et al, 2019;Elshorbagy et al, 2010;Wilby et al, 2003;Nourani et al, 2014;Le et al, 2019;Gauch et al, 2021b). The diversity of modelling approaches reflects the diversity of user objectives, uncertainty in terms of how to best represent the stores and fluxes of water and energy, and the trade-offs in terms of data requirements, degree of realism and computational costs (Beven, 2011).…”

“…DL methods have been used in hydrology and meteorology for decades (Daniell, 1991;Halff et al, 1993;Dawson and Wilby, 1998;Wilby et al, 2003;Peel and McMahon, 2020). However, one architecture explicitly designed for time series simulation, the long short-term memory (LSTM) network (Hochreiter et al, 2001;Hochreiter, 1991), has recently demonstrated credible performance for modelling hydrological signatures across the continental United States (CONUS) (Kratzert et al, 2018(Kratzert et al, , 2019Duan et al, 2020;Feng et al, 2020;Gauch et al, 2021b;Fang et al, 2018Fang et al, , 2020. More recent work has begun not only to explore the accuracy of forecasts but also to use LSTMs to (i) provide estimates of uncertainty (Klotz et al, 2020), (ii) explore the ability of the LSTM to integrate prior physical knowledge into DL model architectures (Hoedt et al, 2021;Jiang et al, 2020), and (iii) to use LSTMs to produce predictions at multiple timescales from a single model (Gauch et al, 2021a).…”

Modeling and interpreting hydrological responses of sustainable urban drainage systems with explainable machine learning methods

“…The source code used in this study is available at https://doi.org/10.5281/zenodo.5652719 (stsfk, 2021), which contains functions for data pre-processing, modeling, modeling result analysis, and plotting. The following R packages are used for modeling and analysis in this research: xgboost (Chen and He, 2020), tidymodels (Kuhn and Wickham, 2020), lubridate (Grolemund and Wickham, 2011), RcppRoll (Ushey, 2018), zeallot (Teetor, 2018), mlrMBO (Bischl et al, 2017), and hydroGOF (Zambrano-Bigiarini, 2020). The following Python packages are used: shap (Lundberg and Lee, 2017), NumPy (Harris et al, 2020), and xgboost (Chen and Guestrin, 2016) Author contributions.…”

Discharge Estimation Using Integrated Satellite Data and Hybrid Model in the Midstream Yangtze River

“…For example, Cheng et al [46] established three machine learning models involving Artificial Neural Network (ANN), Support Vector Regression, and Long Short-Term Memory (LSTM) to predict discharge fluctuation in North China, and satisfactory accuracy was obtained. Gauch et al [47] applied the multi-timescale LSTM model to 516 basins over the continental United States and achieved a significantly better runoff simulation performance than the US National Water Model. Machine learning models have been used to discharge modeling with remarkable success [48].…”