Using a Deep Learning Framework to Forecast Reservoir Water Availability in India
This paper introduces a machine learning-based model to forecast reservoir water volumes in India. In areas with high water stress, having access to timely information on forecasted water availability could help decision-makers avoid the risk of acute water-driven power outages and advocate for long-term, water-prudent policies and management. This forecast can flag when drought-like conditions threaten water supply, but should not be used to inform reservoir management operations.
India is home to the world’s largest population. It is also one of the most water-stressed countries. The increasing demand for water resources, coupled with changes in supply patterns caused by climate change, pose serious implications for the future health and well-being of India’s people and economy. One often overlooked consequence of drought is the risk of power outages due to insufficient water supply. Having access to current and forecasted water availability in reservoirs that supply the power sector with water could help illuminate the chronic issues around India’s dependence on water for power and avoid acute instances of power outages due to insufficient cooling water supply.
WRI introduces a near-real-time forecast of reservoir water volumes for the coming 90 days, using a deep learning framework called the Bayesian long short-term memory sequence-to-sequence-to-sequence model. We show that it is possible to create a high-quality, timely reservoir forecast using global meteorological data. This pilot project provided open access to near-real-time alerts of potential water shortages for 11 reservoirs in India, all of which source water for thermal coal power generation. On average, the 11 pilot reservoirs in this study achieved a coefficient of determination score (which measures how well observed outcomes are replicated by the model) of 92% for a short-term (1-14 day) forecast and 56% for a long-term (15-90 day) forecast. This forecast can be used for a variety of applications, including power production, food security, urban water supply and resilience building.
This publication is a part of WRI India’s Water4Power project, which aims to create a collaborative, open-access platform that guides policymakers, government leaders and civil society organizations in the transition from water-intensive to water-prudent power production. Water4Power is working to make water-related forecasts publicly available to better understand power plant-specific water risks. Additionally, WRI recently partnered with Microsoft, BlackRock and the Growald Climate Fund on the Wave2Wave hackathon, which convened teams from around the world to develop new artificial intelligence predictive models that forecast reservoir water levels in India.
Resources from WRI India:
- Wave2Web: H2Ox (Github)
- Wave2Web Hackathon Overview
- RELEASE: WRI India announces the winners of the Wave2Web Hackathon (October 2021)
- Urban Blues and Prudence: 5 Things to Consider prior to Restoration and Conservation Efforts (March 2020)
- From Waste to Watts: How Sewage Could Help Fix India's Water, Energy and Sanitation Woes (March 2017)
Preview image by Vikramdeep Sidhu/Flickr