Articles | Volume 26, issue 21
Hydrol. Earth Syst. Sci., 26, 5449–5472, 2022
https://doi.org/10.5194/hess-26-5449-2022
Hydrol. Earth Syst. Sci., 26, 5449–5472, 2022
https://doi.org/10.5194/hess-26-5449-2022
Research article
01 Nov 2022
Research article | 01 Nov 2022

Using a long short-term memory (LSTM) neural network to boost river streamflow forecasts over the western United States

Kieran M. R. Hunt et al.

Data sets

River discharge and related forecasted data from the Global Flood Awareness System, v2.1 E. Zsoter, S. Harrigan, G. Wetterhall, P. Salamon, and C. Prudhomme https://doi.org/10.24381/cds.a4fdd6b9

River discharge and related forecasted data from the Global Flood Awareness System, v2.1 E. Zsoter, S. Harrigan, G. Wetterhall, P. Salamon, and C. Prudhomme https://doi.org/10.24381/cds.ff1aef77

River discharge and related forecasted data from the Global Flood Awareness System, v3.1 E. Zsoter, S. Harrigan, C. Barnard, G. Wetterhall, I. Ferrario, C. Mazzetti, L. Alfieri, P. Salamon, and C. Prudhomme https://doi.org/10.24381/cds.ff1aef77

Model code and software

Code for building, testing, and training the models of US LSTM/GloFAS streamflow paper K. M. R. Hunt and G. R. Matthews https://doi.org/10.5281/zenodo.7260860

Download
Short summary
In this study, we use three models to forecast river streamflow operationally for 13 months (September 2020 to October 2021) at 10 gauges in the western US. The first model is a state-of-the-art physics-based streamflow model (GloFAS). The second applies a bias-correction technique to GloFAS. The third is a type of neural network (an LSTM). We find that all three are capable of producing skilful forecasts but that the LSTM performs the best, with skilful 5 d forecasts at nine stations.