Catching the Sun: Using Machine Learning to Predict Solar Panel Output Using Weather
Abstract:
Solar power forecasting is an important aspect of the rise in solar energy. Energy suppliers use it to predict energy supply and determine which assets in their power generation fleet they use. This project explores the use of different machine learning methods to predict solar power generation using weather. It uses data collected by Ph.D. candidate Alexandra Constantin which records the weather and the power output from a solar panel system in Berkley, California from 2008 to 2009. The methods used are a neural network and a boosted trees model. These are then compared to an OLS model which is used as a baseline. The data was standardized and split into testing and training sets. These models were then fitted onto the training set and used the testing set for model validation. The boosted trees model performed the best but is limited in its utility. The neural network performed better than the linear regression but requires more work to meet industry standards.