Forecasting Photovoltaic Performance: A Comparative Assessment of Machine Learning Methods

The increasing global need for renewable energy sources has identified photovoltaic systems as essential to clean energy transition initiatives. This study examines the predictive reliability of machine learning models in assessing and forecasting the output and thermal performance of a medium-scale photovoltaic facility, employing both monofacial and bifacial modules in Canada. A simulation model was created utilizing the System Advisor Model (SAM) with five years of weather data to produce hourly outputs, including power and photovoltaic cell temperature. The datasets were analyzed utilizing multiple regression-based machine learning algorithms, such as Linear Regression, Polynomial Regression, Decision Tree, Random Forest, XGBoost, and regularization methods. Key performance indicators, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R2), were assessed to compare model accuracy. Results demonstrate substantial enhancements in prediction accuracy through the utilization of ensemble models such as Random Forest and XGBoost, attaining R2 values over 97% for both power output and cell temperature forecasts. Bifacial systems exhibited superior energy generation efficiency and tolerance to thermal fluctuations relative to monofacial systems, owing to their dual-sided light capture capability. Analysis of feature importance indicated that Global Horizontal Irradiance (GHI), temperature, and wind speed were the primary determinants affecting performance.