Improved Forecasts of Incoming Solar Radiation Using Machine Learning and Ensemble Weather Model Output

Energy planners must ensure that expected demand on the electric grid is balanced by available supply. Renewable resources, including solar, wind, and hydro power make up an increasing amount of the supply to the grid. However, renewable generation is a function of the weather. This aspect of renewables makes their electric supply highly variable and complicates grid management. For solar energy, incoming solar radiation at the ground surface is impacted by the presence of clouds, aerosols, and shading. Numerical weather prediction models such as the Weather Research and Forecasting (WRF) model provide one of the primary tools used to estimate future solar generation. On top of weather model output, statistical post-processing methods including machine learning are often used to reduce forecast errors. This work uses output from a 15-member WRF ensemble combined with deep learning techniques to improve predictions of incoming solar radiation at surface stations in Vermont. Ensemble forecasts of incoming solar radiation, 2-m temperature, 2-m dew point temperature, 2-m relative humidity, hourly precipitation, 10-m wind speed, and relative humidity at 850, 700, and 500 mb are used with historical observations of incoming solar radiation to train a deep learning model with the goal of improving day-ahead predictions of incoming solar radiation. Our results show employing deep learning techniques on the ensemble output results in up to 30% improvement in forecast accuracy of incoming solar radiation over that of the equivalent raw ensemble forecasts. A discussion on the impacts of including each aforementioned meteorological variable on the accuracy of the deep learning model is also provided.