Short-term Memory Solar Energy Forecasting

Short-term Memory Solar Energy Forecasting at University of Illinois 

Adele Kuzmiakova ”akuzmiakova”, Gael Colas ”colasg”, Alex McKeehan ”mckeehan” December 2017 

Abstract: 

Climate change and energy crisis have motivated the use and development of solar power generation. Since solar power generation is highly intermittent and dependent on local weather characteristics, we apply a variety of linear and non-linear ensemble machine learning models to predict short-term solar energy output at the University of Illinois campus. Implemented models include weighted linear regression with and without dimension reduction, boosting regression trees, and artificial neural networks with and without vanishing temporal gradient. Additionally, we apply a variety of variable selection techniques, which suggest that air temperature, relative humidity, and dew point are the most informative weather parameters for predicting the solar energy generation at the Illinois power plant. Findings from our paper indicate that deep learning with vanishing time-series gradient (long short-term memory layer; LSTM) is the most promising technique for solar predictions based on temporal weather characteristics. The LSTM layer reduces both median training and test error by an order of magnitude compared to the remaining algorithms. This conclusion confirms the presence of short-term temporal dependence in both weather and solar power magnitudes.