Hourly Solar Radiation Forecasting Through Model Averaged Neural Networks and Alternating Model Trees

The objective of the current study was to develop a solar radiation forecasting model capable of determining the specific times during a given day that solar panels could be relied upon to produce energy in sufficient quantities to meet the demand of the energy provider, Southern Company. Model averaged neural networks (MANN) and alternating model trees (AMT) were constructed to forecast solar radiation an hour into the future, given 2003–2012 solar radiation data from the Griffin, GA weather station for training and 2013 data for testing. Generalized linear models (GLM), random forests, and multilayer perceptron (MLP) were developed, in order to assess the relative performance improvement attained by the MANN and AMT models. In addition, a literature review of the most prominent hourly solar radiation models was performed and normalized root mean square error was calculated for each, for comparison with the MANN and AMT models. The results demonstrate that MANN and AMT models outperform or parallel the highest performing forecasting models within the literature. MANN and AMT are thus promising time series forecasting models that may be further improved by combining these models into an ensemble.