New approaches to photovoltaic power estimation

With developing technology, a growing population, and growing industries, the demand for electrical energy is rapidly increasing worldwide. The depletion of fossil fuels and their negative impact on the environment have led humanity to turn to renewable energy sources to address the demand for energy. Among renewable energy sources, solar energy is clean, free, and accessible, making it a popular energy source today. Despite all these positive features, solar energy, by its inherent stochastic nature, causes uncertainty, sudden fluctuations, and stability issues in grid management. Therefore, the widespread use of solar energy necessitates the need for advanced forecasting models to manage this energy. In this thesis, an innovative prediction model that addresses both deterministic and stochastic properties of solar energy data was developed to improve the accuracy of solar irradiance and photovoltaic output power estimations. The study primarily addresses the problem experienced by the Mycielski model with consecutive zero values, and an innovative three-stage method was developed to overcome this problem. Solar irradiance exhibits components influenced by stochastic factors originating from random atmospheric events, as well as deterministic factors such as solar geometry. To address the deterministic component, we used region-specific extraterrestrial solar radiation data. Thus, the developed model provides more accurate predictions of sudden changes in nighttime, sunrise, and sunset times. For the stochastic component, we improved the similar pattern-based Mycielski model to capture random fluctuations, and we also solved the consecutive zero data problem experienced by the Mycielski model. In the second phase of the thesis, an innovative multidimensional Mycielski-based method was developed using the improved Mycielski model to more accurately predict the output power of photovoltaic panels. This approach leverages the Mycielski algorithm's ability to identify similar patterns in multidimensional historical time series data, utilizing the most similar historical sequences for improved prediction. Furthermore, the method was hybridized with AI-based forecasting models and integrated into advanced models such as deterministic-stochastic deep learning architectures. In this multidimensional framework, both electrical quantities and meteorological data were collected through an experimental system established at Afyon Kocatepe University. Experimental results demonstrate that the three-segment Mycielski-based approach developed within the scope of this thesis provides superior forecasting performance compared to traditional models in the literature and avoids the problem of consecutive zero data structures. Furthermore, the hybrid integration of the Multidimensional Mycielski method developed for multidimensional time series with AI-based algorithms successfully improves the prediction of photovoltaic output power, providing a reliable and accurate tool for managing and optimizing solar energy systems.