5.1 NCAR's Gridded Atmospheric Forecast System (GRAFS)

Tuesday, 9 January 2018: 8:30 AM
Room 7 (ACC) (Austin, Texas)
Sue Ellen Haupt, NCAR, Boulder, CO; and D. J. Gagne II, J. Cowie, S. Linden, G. Wiener, and J. A. Lee

Although numerical weather prediction model forecasts have been steadily improving, using artificial intelligence and statistical learning methods can improve those forecasts. Here we describe NCAR’s Gridded Atmospheric Forecast System (GRAFS) and its potential uses. The goal of this effort is to produce a gridded forecast product that blends information from multiple numerical models with smart artificial intelligence techniques and optimizes it based on historical observations. The plan is to make this product widely accessible and applicable for a broad community, working toward joint development and community OpenSource software. GRAFS is modular and customizable, allowing a variety of input sources, output formats, and blending and optimization algorithms.

One application that has shown a need for gridded forecasts, as opposed to point-based forecasts, is distributed photovoltaic power forecasting. Thus, one of the first variables that was explored in GRAFS was global horizontal irradiance. For distributed solar power applications, it is more appropriate to forecast irradiance on a grid, tuned to available observations, even if those observations do not match the precise locations of the PV arrays. Subsequently, GRAFS has been expanded to include gridded optimized forecasts of temperature, dewpoint, and wind speed.

GRAFS has a modular, customizable framework that simplifies adding new uses, allowing a variety of input sources, output formats, and blending and optimization algorithms. Techniques include employing the Dynamical Integrated Forecast System (DICast®) as a blending tool, as well as using artificial intelligence methods including random forests, regression trees, and gradient boosted trees. Several methods to smooth the forecast have been tested. Plans to include more variables, to test additional methods, and to embed statistical assessment methods based on the Model Evaluation Tools are planned for the near future.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner