Machine learning for solar irradiance forecasting

Machine learning techniques were investigated to forecast surface-level solar irradiance and, by proxy, the output of a solar farm near Athens, GA, for 1-24 hours into the future. Weather predictions from the National Oceanic and Atmospheric Administration (NOAA) and ground-based weather observations from the Georgia Automated Environmental Monitoring Network (GAEMN) served as inputs to the models. Various learning algorithms were compared, and an analysis of the relative importance of potential model inputs was performed.Also examined was the ability of machine learning models to generalize to unfamiliar weather and climate conditions. In the experiments, random forests outperform several other techniques commonly found in the literature by as muchas 26.8%. The random forest models benefit from incorporating weather forecasts from a grid of cells surrounding the target site, resulting in improvements to accuracy up to 4.1% relative to models that do not make use of such information.