Solar power generation forecasting using machine learning techniques

Abstract

The use of solar energy introduces challenges due to its variability over time, making accurate forecasting essential for a reliable integration into power system operations. Machine learning models have become a key tool in energy systems, enabling data-driven approaches to forecast complex and nonlinear phenomena such as renewable power generation. This thesis evaluates and compares three machine learning models—Random Forest, LightGBM, and Long Short-Term Memory (LSTM)—for short-term solar power forecasting with a 3-hour ahead prediction horizon. A 3-hour ahead horizon is particularly relevant for operational planning and short-term energy management in power systems.

The study uses hourly photovoltaic (PV) data from the PVGIS tool for Madrid, Spain, covering the period 2016–2020. The data were preprocessed and used to train the models on 2016–2019 data and evaluate them on 2020 data. Model performance was assessed using Root Mean Square Error (RMSE), coefficient of determination (R²), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE).

The results showed that Random Forest achieved the best overall performance with an MAE of 39.46 ± 0.03 kW and a MAPE of 26.93 ± 0.05%, while LightGBM achieved the lowest RMSE of 89.00 ± 0.11 kW and the highest R² of 0.8777 ± 0.0003. The LSTM model showed lower accuracy for all metrics, with an MAE of 45.87 ± 1.88 kW.

These findings indicate that tree-based machine learning models are more suitable for short-term solar power forecasting with the given dataset and offer a practical approach for operational energy management.