A Weight-Varying Ensemble Method for Short-term Forecasting PV Power Output.

Abstract The photovoltaic (PV) power generation has randomicity and intermittence due to the influence by meteorological factors and it is of significance to establish a reasonable prediction model for PV power. Since it is tough to get a completely optimal prediction of PV power by a single model due to the existence of uncertainty, the paper considered proposing an ensemble method by integrating various individual models to achieve better accuracy. In this paper, a weight-varying ensemble (WVE) forecasting model is established to improve the precision of the short-term prediction of PV power. First, the extraction of feature vectors was implemented to find out the most important variables which also helps to improve the accuracy through data pre-processing. Second, the weather pattern was recognized and clustered based on a self-organizing feature map (SOM) method. Third, the five min-ahead prediction output was obtained using the WVE model that utilizes an integrated framework of Generalized Regression Neural Network (GRNN), Extreme Learning Machine Neural Network (ELMNN) and Elman Neural Network (ElmanNN), which were assembled by Genetic Algorithms optimized Back Propagation Neural Network (GA-BPNN). Results show that the WVE model achieved a higher accuracy with mean absolute percentage error (MAPE) of 5.17%, 5.26%, 5.49% and 5.82% for four types of data. The GA can be applied to optimization of other weight-varying ensemble problems. [ABSTRACT FROM AUTHOR]