SBIR-STTR Award

Hydropower operators rely on forecasts for planning and operational decision-making, but the inaccuracy of current forecasts has implications for revenue, risk, and regulatory requirements. Our proposal leverages machine learning and satellite imagery innovations to improve the accuracy and reliability of hydrologic forecasts. Using a distributed machine learning model, we are able to efficiently incorporate satellite earth observations, model uncertainty, and enable hydropower operators to better plan and manage their assets while reducing risk. In addition to the hydropower utility market, we have had early validation that our approach has application for conservation NGOs, source water utilities, government agencies, and agriculture - illustrating the long-term commercialization potential of our project for a wide range of stakeholders. Our project will address the following central challenges for existing forecasting methods: Existing models underperform compared with machine learning prototypes, existing models do not take advantage of all available satellite earth observations, and existing models do not take advantage of all available dynamic weather data. To address the challenges, in Phase I of our project we will incorporate satellite data into a spatially distributed neural network model, providing a detailed understanding of current land surface conditions and enabling the model to produce more accurate forecasts., account for uncertainty and variability in weather forecasts by integrating ensemble weather forecasts from multiple weather forecasts to create a more accurate probabilistic prediction, perform a large scale forecast verification and benchmarking exercise to demonstrate the value provided by satellite data driven machine learning hydrology forecasts. Ultimately, the resulting technology will be used to inform reservoir operations optimization and enable hydropower operators to maximize revenue and reduce risk, while supporting environmental aims and meeting regulatory requirements.

Existing approaches for predicting streamflow on shortterm 110 day and seasonal 1090 day horizons are not sufficiently accurate, flexible, or scalable to inform smart and flexible water decisions across the planet. This problem is clear today in the hydropower industry, for water allocation managers, and is becoming more acute for conservation organizations as ecosystem protection grows in complexity. These groups rely on forecasts for planning and operational decisionmaking, but shortcomings in current forecasts hamper decision makers’ ability to plan operations, prepare for extreme weather events and climatic shifts, and protect ecological services. The overall objective of the Phase I and proposed Phase II project is to advance the technical and commercialization of a novel hydrologicallyguided machine learning approach to predict streamflow. Through integration of near realtime satellite basin data, weather forecasts, and ontheground observed data, the model delivers forecasts with demonstrated accuracy using a cloudbased system that is scalable, resilient, and secure. In Phase I, an initial prototype was transformed into a fully operational system and commercialized for a cohort of initial customers. The three main accomplishments from Phase I were: 1 built, tested, and implemented a distributed neural network model that creates accurate forecasts in large> 20,000 mi2 basins; 2 demonstrated the added value in accuracy that results from including multiple weather forecast sources as parallel model inputs; and 3 operationalized an evaluation system such that experimental model versions can be easily compared with competitors’ models and alternative methods. During Phase I, operational forecasts were deployed at 36 points. The Phase II objectives build on findings from Phase I, and broadly aim to: 1 increase accuracy by including additional weather forecast sources that span the shortterm and seasonal products, 2 quantify the model’s financial benefits over alternative approaches to communicate the return on investment with target customers, and 3 build out commercial scalability of the model system and delivery methods. Achievement of these objectives will enable and accompany a concerted business effort to scale from the initial cohort of customers by integrating results from 2 and 3 into the product marketing strategy. In the shortterm, three user groups will benefit from the technology: hydropower entities will optimize generation and operate safely, helping transition the grid to renewable sources while reducing their environmental impact and increasing revenues; water regulators will make datainformed allocation decisions; and conservation organizations will prioritize restoration projects more effectively.