NREL and IBM Improve Solar Forecasting with Big Data
IBM, ESIF researchers, and a team of partners built a better solar forecasting model using deep-machine-learning technology. The multi-scale, multi-model tool, named Watt-sun, learns from past predictions to continuously improve its solar forecast accuracy.
The complexity of solar forecasting means that even the best models have weaknesses that leave utilities with costly uncertainties. To get the full value out of solar technologies, this approach blends multiple models and forecast data together to identify what mix works best to predict solar resources.
One technical challenge that needed to be addressed to build this complex and data-intense multi-model was to effectively quantify forecast accuracy. With so many factors at play (location, cloud cover, time of day), no consistent set of metrics existed to measure the accuracy of a solar forecast. As part of this project, NREL led the development of the first standard suite of metrics for this purpose.
Validating Watt-sun at multiple sites across the United States demonstrated a more than 30% improvement compared to forecasts based on the best individual model and a more than 10% improvement compared to conventional machine-learning forecasts.
This project was developed through the U.S. Department of Energy’s SunShot Initiative.