Solar Radiation Prediction Improvement Using Weather Forecasts

Prediction models were developed to generate forecasts of solar radiation, and, by proxy, expected solar plant power output, for one hour and 24 hours in the future. Data was sourced from the Georgia Automated Environmental Monitoring Network (GAEMN) and the National Oceanic and Atmospheric Administration (NOAA) for five cities in Georgia. Early predictive models only made use of historical recorded solar radiation and other weather phenomena as inputs, while later models incorporated weather forecasts for the target area and surrounding areas. Including weather forecast data in the prediction models resulted in a 7.6% reduction in mean absolute error (MAE) for one-hour predictions when compared to using historical observations alone, and a 40.2% reduction in MAE for 24-hour predictions. Results from several machine learning techniques were compared, with Random Forests achieving the lowest error rate. The results indicate that weather forecasts are an important component of accurate solar radiation prediction even over short- and medium-term prediction timeframes, and the inclusion of the surrounding geographical area in addition to the target city is an important component of these predictions.

I. Introduction

Solar power is unlike other forms of energy production such as nuclear power or fossil fuels in that it relies heavily on an external agent as the key determiner of power output. Output is discernibly higher on sunny days during the summer, and lesser on cloudy days or during the winter when the sunlight hits the earth at a less direct angle. Making accurate predictions about expected solar power output is paramount to efficiently harvesting power from a solar plant.