Wave2Web: Near-real-time reservoir availability prediction for water security in India

This is a Preprint and has not been peer reviewed. This is version 1 of this Preprint.

Add a Comment

You must log in to post a comment.


There are no comments or no comments have been made public for this article.


Download Preprint


Lucas Kruitwagen , Chris Arderne, Thomas Lees, Lisa Thalheimer, Samantha Kuzma, Samrat Basak


By 2050, over half the world's population will live in water-stressed areas. Medium-term drought forecasting can help planners avoid ``day-zero'' events and adapt to climate change. Machine learning-based precipitation-runoff modelling enables the prediction of surface water flow using only the meteorological record of a water basin.

In this work, we extend a Bayesian LSTM precipitation-runoff model with graph convolutions based on hydrological basin adjacency to predict reservoir water availability for sixty-six reservoirs in India. Employing a "sequence-to-sequence-to-sequence" model allows predictions to be based on the combination of meteorological forcing data and ex-ante forecast data while producing predictions to a ninety-day future horizon. On a held-back test set of daily reservoir water availability changes, we achieve a coefficient of determination of 0.372 for the maximum likelihood estimate for the 1-to-5 day horizon averaged across all sites, which reduces to 0.337 for the 75-to-90 day horizon. We also find that removing the graph convolutional layer increases mean performance by 0.82 percentage points over the same horizons. This work represents the winning submission of the Wave2Web hackathon; the code and data is publicly available and near-real-time predictions are available at h2ox.org.




Planetary Hydrology


hydrology, Deep learning, LSTM, graph convolution


Published: 2022-06-02 05:53

Last Updated: 2022-06-02 12:53


CC BY Attribution 4.0 International

Additional Metadata

Conflict of interest statement:

Data Availability (Reason not available):
All data used in this work is obtained from publicly available sources. The dataframes used to train each basin network are publicly available at https://console.cloud.google.com/storage/browser/oxeo-public/wave2web/h2ox-ai. The trained models are available via github at https://github.com/H2Oxford/h2ox-ai/tree/main/models. Model inference is provided in near-real-time via the api.h2ox.org data service. Documentation for this service can be found at https://api.h2ox.org/docs#/. Access to intermediary BigQuery tables and Cloud Storage Zarr archives can be made available on request.