Improved Forecasting Models Will Help Grid Operators Predict Production of Solar Power  

WASHINGTON D.C. – This week, during a live, virtual event at Carnegie Mellon University’s Energy Week, the U.S. Department of Energy (DOE) announced the winners of the American-Made Solar Forecasting Prize. The five winning teams, who will each receive $50,000, had the best-performing forecasting models and strongest plans to accelerate the adoption of probabilistic forecasts. The two runners-up will receive $25,000 each.  

The Solar Forecasting Prize is designed to incentivize the development of state-of-the-art solar forecasting capabilities to assist grid operators in predicting how much solar power will be produced while considering weather-related uncertainties such as cloud coverage.  

Accurate solar forecasting is essential for reliable and cost-effective integration of solar energy into the electricity grid, helping to achieve the Biden Administration’s goal of decarbonizing the grid by 2035. 

“As the United States deploy more solar energy on the electric grid, accurate forecasts will be key for grid operators to maximize the potential of this technology,” said Kelly Speakes-Backman, Principal Deputy Assistant Secretary for the Office of Energy Efficiency and Renewable Energy at DOE. “Through their models, these winning teams have advanced forecasters’ ability to predict the amount of power that solar energy generators could reliably deliver in a given 24-48 hours.” 

Every day for four weeks, teams submitted day-ahead solar forecasts for predetermined locations. Competitors also submitted plans that described commercialization approaches for their forecasting tools or innovative ideas in order to accelerate the adoption of advanced forecasts by system operators, integrated utilities, and other users.  

The winners are:  

  • Nimbus, AI (Honolulu, Hawaii)   
    Fast Solar Forecasting with Machine Learning
    This team developed a fast and inexpensive system for geographically flexible, hyper-local day-ahead probabilistic solar forecasting. The team combined historical ground- and satellite-based instrument data with physics-based numerical weather prediction (NWP) techniques to produce probabilistic forecasts.
  • University of Michigan CLaSP Solar Forecast Team (Ann Arbor, Michigan)  
    A Novel Hybrid Approach for Solar Forecasting
    This team developed a hybrid solar forecasting method using ground horizontal irradiance based on a recursive neural network (RNN), trained with past observations and a weather-regime-dependent empirical bias correction scheme. The team based this scheme on the RNN output and the multi-day weather forecast made from NWP. 
  • Northview Weather (Danville, Vermont)  
    Determining Who Wins the Cumulus vs. Stratus Battle
    This team developed a solar forecasting method that uses a mesoscale weather forecast model to produce a dynamically based spread of probabilistic solar power forecast information. Probabilistic information is also tuned blending statistical information and machine learning of historic sky cover observations. 
  • WenYuan Tang (Apex, North Carolina)  
    A Hybrid Approach to Probabilistic Forecasting
    This team developed a simple (low-data-cost and low-computational-cost) yet effective solar forecasting method that includes learning from many base models, as well as physical and statistical models, to provide a comprehensive tool for both forecasting and grid optimization. Their simple yet effective model is easy to interpret, train, validate, and deploy, thus overcoming the barriers of experimentation and adoption by utilities and other end users. 
  • Leaptran (San Antonio, Texas)  
    Integrated Solar Forecasting Solutions
    This team used crowd-sourced weather data and algorithms, leveraging site-specific data fusion, to achieve intra-day and days-ahead solar forecasting. This model offered a >50% improvement of days-ahead, intra-hour, and intra-day solar forecasting accuracy by integrating asset-level data. 

The runners-up are:  

  • Matt Motoki (Aiea, Hawaii)  

    This team developed an advanced machine learning technique for probabilistic forecasting that utilizes custom deep neural networks to directly minimize the Continuous Ranked Probability Score loss with no post-processing calibration needed. This method allowed for greater accuracy and its lightweight architecture allows it to run faster than NWP ensembles and other machine learning approaches. 
  • Syracuse University Team (Syracuse, New York)  
    Weather Adaptive Probabilistic Solar Forecast
    This team developed a weather-adaptive probabilistic day-ahead solar forecast methodology that leverages innovations in machine learning and statistics. This model could adapt to different weather patterns for an accurate forecast under varying weather conditions. 

The prize featured the use of an assessment framework, that allows for transparent and consistent analysis and evaluation of solar forecasts, called the Solar Forecast Arbiter (SFA). The SFA is an open-source platform that was developed by the University of Arizona with funding from the DOE Solar Energy Technologies Office (SETO) in 2018. The teams used the SFA to compare their forecasts against the platform's benchmark for quantitative evaluation of their forecast accuracy. The prize administrator relied on the SFA results for the evaluation of each team’s final standing.  

The American-Made Solar Forecasting Prize was administered by the National Renewable Energy Laboratory and funded by SETO. Learn more about DOE’s research to integrate solar into day-to-day electricity system operations