Abstract
The integration of Solar Energy in smart grids and many utilities is continuously increasing due to its environmental and economical benefits. However, the uncertainty of available solar energy provides challenges regarding stability in power generation and therefore consistency in production-consumption balance. The availability of an efficient solar energy short-term predictor is found adequate to remedy such type of problems, and therefore permits to enhance the overall system reliability and power generation scheduling. In this paper we survey several ensemble methods based on machine learning named stacking for one-day-ahead predicting of solar radiation intensity. The support vectors regressors are used as base forecasters in an ensemble, and the Multi-layer perceptrons as well as the Decision Trees and K-Nearest Neighbor Regressors are used as meta-learners to combine the predictions of the ensemble in a stacking mode. The sliding window technique is used in a preprocessing phase to reframe the time series predicting problem as a supervised learning problem. The combiners provide a smart weighted averaging of the forecasters' outputs. The solar radiation prediction of the surveyed models has been evaluated and analyzed over an entire year. The stacking based ensemble models outperformed all individual models as well as other combining techniques.

This publication has 21 references indexed in Scilit: