Forecasting of solar radiation with intelligent hybrid approaches

Nowadays, studies on new and alternative energy resources are of great importance in order to meet the energy needs of future generations due to the decrease in fossil-based energy resources that are rapidly depleting. Renewable energy sources constitute a large part of the studies on such energy sources. Solar energy, which has a very important place among these resources, is an indispensable resource that can be found at any point in the world, abundant and inexhaustible and free from polluting wastes. Considering the disadvantages of the energy to be obtained from solar energy systems such as not continuous and having high costs, feasibility studies are very important before the installation of solar power plants. In these feasibility studies to determine the potential of solar energy, the amount of solar radiation intensity appears as important parameters. Many different methods such as statistical, physical, machine learning and hybrid approaches in the estimation of solar radiation intensity have been proposed in the literature. In this study, hybrid models are proposed by using artificial neural networks (ANN), which is a powerful tool in short-term solar radiation intensity estimation, together with both wavelet decomposition method (WD) and empirical mode decomposition method In the hybrid model, first of all, monthly solar radiation intensity data of Bilecik province were separated in four levels by WD and EMD method. In the next stage, the model results were obtained by training with a feed forward network with historical time data for each level. The test performance results of each hybrid network structure created are compared with the outputs obtained using only ANN. It was observed that the FFNN_EMD hybrid approach gave results with higher accuracy according to the error performance metrics of the results obtained in the monthly basis. Especially with the rapidly developing machine learning techniques, it is predicted that such hybrid approaches can be applied to different intelligent heuristic methods in the future.