A Deep Learning Architecture for Conservative Dynamical Systems: Application to Rainfall-Runoff Modeling

Spotlight talk at the AI for Earth Sciences workshop of the NeurIPS 2020, presenting a first glimpse of a new LSTM-based model that conserves mass by design: the Mass-Conserving LSTM

Abstract

The most accurate and generalizable rainfall-runoff models produced by the hydrological sciences community to-date are based on deep learning, and in particular, on Long Short Term Memory networks (LSTMs). Although LSTMs have an explicit state space and gates that mimic input-state-output relationships, these models are not based on physical principles. We propose a deep learning architecture that is based on the LSTM and obeys conservation principles. The model is benchmarked on the mass-conservation problem of simulating streamflow.

Paper

Nearing, G. and Kratzert, F. and Klotz, D. and Hoedt, P.-J. and Klambauer, G. and Hochreiter, S. and Gupta, H. and Nevo, S. and Matias, Y.: “A Deep Learning Architecture for Conservative Dynamical Systems: Application to Rainfall-Runoff Modeling”. Workshop on AI for Earth Sciences 34th Conference on Neural Information Process-ing Systems (NeurIPS 2020) Vancouver, Canada.

Video of the presentation: https://slideslive.com/38941248

Citation

@inproceedings{nearing2020neurips,
  title={A Deep Learning Architecture for Conservative Dynamical Systems: Application to Rainfall-Runoff Modeling},
  author={Nearing, G. and Kratzert, F. and Klotz, D. and Hoedt, P.-J. and Klambauer, G. and Hochreiter, S. and Gupta, H. and Nevo, S. and Matias, Y.},
  booktitle={Workshop on AI for Earth Sciences 34th Conference on Neural Information Process-ing Systems (NeurIPS 2020)},
  venue={Vancouver, Canada},
  date={6-12 Dec},
  year={2020}
}