Estimation of solar panel energy production with different machine learning methods

Solar energy is increasingly playing a crucial role in the global energy transition as a sustainable and environmentally friendly energy source. However, the natural variability in solar irradiance makes reliable forecasting of energy production a critical necessity. Traditional statistical methods often fall short in this regard; therefore, artificial intelligence-based approaches are gaining prominence. This thesis aims to accurately predict solar energy generation using hourly production and meteorological data from 12 different solar power plants in Türkiye, covering the years 2023–2024. The forecasting process involved data preprocessing followed by the implementation of various machine learning algorithms and an advanced deep learning model incorporating attention mechanisms, known as the Temporal Fusion Transformer (TFT). Machine learning models were optimized using four different hyperparameter tuning techniques. Model performances were evaluated using R² (coefficient of determination), MAE (mean absolute error), MAPE (mean absolute percentage error), and RMSE (root mean square error) metrics. According to the results, the Random Forest and Extreme Gradient Boosting algorithms achieved high accuracy, with approximately 96.20% R² and a MAE of around 113. Another key finding of the study is that the Temporal Fusion Transformer, which is effective in modeling complex relationships in time-dependent data, outperformed all other models with 96.81% R² and a MAE of 76.35 on the test dataset. These findings demonstrate that deep learning-based advanced approaches, especially models specifically designed for time series data, are highly effective in solar energy forecasting and can significantly contribute to planning, management, and decision-support processes in the energy sector.