A novel clustering approach for short-term solar radiation forecasting

Abstract

This paper proposes a hybrid solar radiation forecasting method based on a novel game theoretic self-organizing map (GTSOM). New strategies are proposed to resolve the limitations of the original SOM for non-winning neurons and increase their competition with winning neurons to obtain more input patterns. Neural gas (NG) and competitive Hebbian Learning (CHL) are used to enhance the learning and quality of the map. Solar radiation data are decomposed by the discrete wavelet transform (DWT). A time series analysis is then used to develop the structure of the training and testing for Bayesian neural networks (BNNs). The proposed GTSOM groups the time-series analyzed datasets into clusters with similar data. The elbow method is used to determine the number of clusters. A cluster selection method is developed to determine the appropriate cluster whose solar radiation data provide the input to the NN. Temperature, wind speed and wind direction data are also included in the inputs to the BNN whose outputs provide the solar radiation forecasts. The historical solar radiation data are used to evaluate the accuracy of the hybrid forecasting with the proposed clustering and its comparison with that of the K-means, the original SOM and NG clustering algorithms. The comparison demonstrates the superior performance of the proposed clustering method.