Estimation of Solar Radiation with Artificial Neural Networks

Abstract

Accurate estiınation of solar radiation is of great importance for the effective use of solar energy and effıcient ınanagement of energy grids. This study evaluated the effectiveness of Artificial Neural Networks (ANN), especially the NARX (Nonlinear Autoregressive with Exogenous lnputs) model in solar radiation estiınation. The analyses conducted for nine different cities in Türkiye examined the estiınation perforınances of the cities during the ınachine leaming processes, including the leaming, validation, and testing stages. According to the analysis results, Antalya and Mersin cities exhibited the best perforınance with the lowest Mean Square Error (MSE) values (0.2931 and 0.2897) and high correlation coefficients (R=0.9799 and R=0.9797). Trabzon had the highest MSE values (0.6203) and performed less than other cities. Despite relatively high error rates, Istanbul City strongly correlated. Evaluating the performance criteria during training revealed that İzmir and Istanbul demonstrated superior performance with low error values. According to the gradient values calculated during the training, lstanbul reflected the smoothest optimization process with the lowest value (33.3). Higher gradient values revealed the data complexity in cities such as Trabzon (84.4) and Van (57). These fındings show that the NARX model is effective in solar radiation estimation and is a powerful method, especially in short-term forecasts. We evaluated the model's consistent performance in different cities by considering data complexity and computational efficiency. The study recommends applications of the NARX method on a larger scale and supported by different meteorological variables for future research.