AI Solar Nowcasting System

About

The global economy is undergoing a rapid transition to a clean energy future, underscored by the acceleration of renewable energy. All segments of power generation, grid management, and consumption must adapt to facilitate this shift. Climate-related risks and sustainability considerations are steadily becoming central to all stakeholders’ environmental commitments. Alongside this transformative landscape, solar power is quickly becoming an integral renewable resource for both utilities and consumers. For utilities, intermittent solar generation throughout the day presents challenges as they may, for example, need to use quick-starting capacity relying on the power inertia of fossil fuels to ensure power resilience, particularly as demand-side behind the meter solar production steadily increases. Forecasting short-term solar patterns is important for system efficiency and reliability and can lead to optimized use of this generation capacity. Traditional solar pattern forecasting either involves on-the-ground hardware deployment which may be capital intensive and unable to keep up with the rapid adoption curve of solar photovoltaic (PV) installation sites (especially within the geographically dynamic residential segment) or satellite intelligence based on physics-based numerical models which are not well-suited to short-term forecasting and are computationally expensive. These shortcomings prevent the widespread adoption—across geographies, energy segments, and economic status—of forecasting systems which are important in facilitating the world’s transition to more renewable energy. Our team of University of Hawaiʻi (UH) researchers has developed a cutting-edge machine learning system for forecasting ground-level solar irradiance and photovoltaic production across short time scales from satellite data, and they have formed Nimbus AI, LLC to launch this technology. This system is based on location-specific convolutional neural networks trained on historical data, which is more computationally-efficient than traditional numerical models. Initial site deployment data indicates a system accuracy comparable with other current state-of-the-art satellite and/or ground-based forecasting systems. The forecasting system and curated data products have been deployed using the US National Oceanographic and Atmospheric Administration’s GOES-17 satellite which sits in geostationary orbit above the equator at 137° West and has a high data refresh cadence of 10-minute intervals. This facilitates a service area for the Polynesian islands of the Pacific, Fiji, the Galapagos Islands, Central America, and the western half of the United States. The Nimbus AI data pipeline can pull images as they are uploaded and produce forecasts for all targeted sites and swaths within seconds. Moving forward, geographies around the world can be similarly targeted by leveraging other satellite data sources as needed. Technological Features Fast, location-agnostic deployment Zero on-the-ground hardware cost. Immediate deployment potential across the Pacific. Near-term deployment to other global geographies based on satellite coverage. Cutting-edge satellite and AI incorporation Leverages next-generation satellite data with 500 square meter resolution from multiple data streams. Leverages state of the art machine learning AI methods for forecasting. Data pipeline efficiency Proprietary in-house data pipeline facilitating rapid, scalable uptake and assimilation of satellite data sources.