Investigation of suitable machine learning methods for power predicton of a photovoltaic plant under different environmental conditions

Renewable energy sources (RES) have gained significant importance in the energy sector due to the limitations of fossil fuel consumption and its environmental impacts. Photovoltaic (PV) systems offer an environmentally friendly option for generating electricity from solar energy; however, their production levels are influenced by weather conditions. This makes accurate forecasting of solar energy production critical for maintaining the balance between supply and demand in energy systems. This thesis, driven by this necessity, aims to predict the output power of PV panels using weather data and machine learning techniques. In this thesis, two different types of Wavelet Transform (WT), Discrete Wavelet Transform (DWT) and Undecimated Wavelet Transform (UWT), have been used for feature extraction. The extracted features were analyzed using regression and tree- based machine learning models. DWT demonstrated the lowest errors and best performance in Support Vector Regression (SVR) with 0.0001 Mean Squared Error (MSE), 0.0025 Root Mean Squared Error (RMSE), and 0.9992 Coefficient of Determination (R2), while it showed the lowest prediction times (0.0003 sec.) in Ridge Regression (RR) and Lasso Regression (LR), making it more compatible with regression-based models. On the other hand, UWT performed better in tree-based methods such as Random Forest (RF) and Decision Tree (DT). In particular, the UWT- DT model provided the most balanced and excellent result with MSE=0.00004, RMSE=0.0058, and R2=0.9997. The study concluded that appropriate feature extraction and model selection, based on the data type and problem structure, have a decisive impact on prediction accuracy. Model performance was compared using evaluation metrics such as MSE, RMSE, R2, prediction, and training times, with tree-based models standing out due to their lower error rates (MSE=0.00004-0.00009). Regression-based models, on the other hand, offered faster prediction times (LR=0.0002). The study was conducted in MATLAB, utilizing real-time data collected from a 5 kW Grid-Connected PV System and a 2 kW battery group at the Mimar Sinan Campus of Bursa Technical University (BTU), used in the "Renewable Energy Supported Electric Vehicle Charging Station" prototype system. The dataset, consisting of environmental factors, was employed to model the PV panel output power in detail under various weather conditions. The use of real- time measurement data significantly enhances the effectiveness of the forecasting models in practical applications. This thesis demonstrates the applicability of optimized feature extraction methods and machine learning models for the prediction and management of renewable energy systems. The results show that UWT and tree-based models perform better in complex structures, while regression-based models offer time efficiency. In this context, selecting models suitable for the data type is crucial in improving the efficiency of energy systems. Future studies are recommended to explore the use of production forecasts in energy storage and electric vehicle charging stations. Keywords: PV power system, Power prediction, Machine learning, Wavelet transform, Regression models, Decision tree