Scaling Machine Learning Models to Produce High Resolution Gridded Solar Power Forecasts

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Monday, 5 January 2015: 4:00 PM
124B (Phoenix Convention Center - West and North Buildings)
David John Gagne, NCAR, Boulder, CO; and S. E. Haupt, S. Linden, J. K. Williams, A. McGovern, G. Wiener, J. A. Lee, and T. C. McCandless

Accurate forecasts of solar power production are a necessary component in making solar a cost-effective and viable renewable energy source. The Department of Energy has funded a major project led by the National Center for Atmospheric Research to produce improved forecasts of solar power by incorporating advances in real-time observations, numerical weather prediction modeling, and machine learning techniques. As part of this project, a gridded forecasting system is being developed to combine observations and predictions from multiple numerical weather prediction models into consensus forecasts of solar irradiance and solar power plant production. Developing this system requires addressing the challenges of processing large amounts of model output, training and applying machine learning models to data at a large number of observation locations, and inferring values at every grid point in the domain by accounting for local and large-scale influences. This presentation discusses the initial methods used to address these challenges. As part of the development process, the accuracy and speed of multiple types of preexisting and custom machine learning models are compared. Varied parameter settings and combinations of models are also examined. Case studies showcasing model performance in challenging situations are also included. Some avenues for future development of the system, and how it may be applied to other gridded forecasting problems, are also explored.