A Comparative Machine Learning Approach to Forecasting Solar Power Across Diverse Climate Conditions رسالة ماجستير

Abstract

The study mainly compares the efficiency of Machine Learning (ML) models in forecasting solar photovoltaic (PV) power generation under different climatic conditions reported between August 2022 and July 2023. The data were collected from two locations: Tubas in Palestine, with highly variable weather, and Balearic Islands in Spain, with stable conditions. The dataset includes meteorological data, irradiance, temperature, humidity, pressure, and wind speed from NASA’s POWER database, and solar power generation data from Tubas Electricity Company and Red Eléctrica de España. The study involved comprehensive data cleaning, preprocessing, exploratory data analysis, and model development. ML models—Linear Regression, Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Bidirectional Long Short-Term Memory (Bi LSTM), and a hybrid Convolutional Neural Network - Long Short-Term Memory (CNN LSTM)—were implemented. Model performance was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²) metrics. Deep learning models, particularly Bi-LSTM and hybrid CNN-LSTM, achieved better performance across both regions. Tubas presented greater forecasting challenges due to fluctuating weather, despite that the deep model remained robust. Key findings showed that solar irradiance was the most influential predictor in both regions, while temperature, humidity, and wind speed significantly contributed under fluctuating conditions. The study concludes that deep learning models are best suited for solar forecasting under diverse weather conditions. The findings are valuable for enhancing solar energy integration, guiding infrastructure investments.