Hydrology

Rainfall–runoff modelling using Long Short-Term Memory networks

A foundational evaluation of sequence learning across 241 catchments.

Research summary2 min readPublished 22 November 2018
Authors
Frederik Kratzert; Daniel Klotz; Claire Brenner; Karsten Schulz; Mathew Herrnegger
Institution
University of Natural Resources and Life Sciences, Vienna
Journal
Hydrology and Earth System Sciences 22, 6005–6022
Geography
United States catchments (CAMELS)
Method
LSTM; hydrological benchmarking; regional transfer
DOI
10.5194/hess-22-6005-2018

Research question

Can a recurrent neural network learn rainfall–runoff behaviour from meteorological inputs and benefit from information across catchments?

Method and findings

Kratzert and colleagues evaluated LSTM models using 241 catchments in the CAMELS dataset. They compared performance with the Sacramento Soil Moisture Accounting model coupled to Snow-17, and examined regional learning and transfer to individual catchments. Their transfer approach improved performance relative to single-catchment LSTM training and the stated hydrological benchmark.

Limitations

The reported experiments establish performance for their data and evaluation setting. They do not by themselves validate extrapolation to a changed climate, a new catchment or a drought operating policy.

Engineering interpretation

Our interpretationFor water-supply planning, evaluate low-flow duration and cumulative volume as well as an overall hydrological score. Carry uncertain inflows through the reservoir operating model. Good streamflow prediction and a reliable supply decision are separate evaluation tasks.

Explore the practical modelling workflow.

Sources & further reading

Source findings are distinguished from editorial interpretation. Apply current local criteria and project evidence when making engineering decisions.