Machine Learning Based Multi-model Blending for Enhancing Renewable Energy Forecasting

With increasing penetration of solar and wind energy in global power grids, the pressing need for accurate forecasting has become well-recognized. In this presentation, we report significant progress towards a scalable solar/wind forecast system using multi-model blending to enhance accuracy. The system leverages upon multiple existing physical models for prediction including numerous atmospheric and cloud prediction models based on sky camera and satellite imagery as well as numerical weather prediction (NWP) products. By regressing historical predictions vs. measurements, one obtains an optimal blending of the individual models to create a “super model” using machine-learning strategies, similar to those used by IBM Watson^TMin the Jeopardy! Grand Challenge.

More importantly , we demonstrate that in addition to parameters (solar/wind power, solar irradiances) to be predicted, including additional atmospheric state parameters which collectively define “weather situation categories” as machine learning input provides further enhanced accuracy for the blended result. Functional analysis of variance is applied to show that the error of individual model typically has substantial dependence on “weather situation categories”. The machine-learning system effectively reduces such situation dependence error thus produces more accurate result compared to conventional multi-model ensemble approaches based on simplistic equally or unequally weighted model averaging. Results using the system show over 30% improvement in solar irradiance/power forecast accuracy compared to forecast based on the best individual model.

The work is partially supported by Department of Energy SunShot Initiative contract #DE-EE0006017.