An evolutionary method for creating ensembles with adaptive size neural networks for predicting hourly solar irradiance

In this paper we propose a hybridized approach for finding high quality artificial neural network (ANN) for calculating hourly estimates of solar irradiance. These properties are essential for performance analysis of solar based energy generation. To be more precise the hourly global horizontal irradiance (GHI), direct normal irradiance (DNI) and diffuse horizontal irradiance (DHI) are estimated based on ANNs which are trained using satellite and ground measurement data. In the proposed method we explore the effect of combining the measured data with properties derived from the standard physical models. The performance of the method is improved by using a genetic algorithm in two ways. First by selecting the parameters that are used for training the ANN. Secondly by adapting the size of the hidden layer of the ANN based on the number of selected input parameters. The adaptive size based approach proves to be especially suitable for ANN ensembles. In our computational experiments we evaluate the effectiveness of the proposed method on feedforward neural network. The results show that the adaptability of the ANN manages to notably improve the performance when compared to the standard approach using a fixed size of the hidden layer.