Machine learning-based estimation of power output in solar photovoltaic systems under real-world conditions

Precise prediction of DC power output from photovoltaic (PV) systems under real conditions is essential for improving efficiency and detecting degradation. This paper presents a machine learning framework to forecast PV power using electrical signals (e.g., voltage and current) and environmental data (irradiance and temperature). Three regression models -eXtreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and Artificial Neural Networks (ANN) - were trained on 13,923 samples from a 455.4 kWp solar PV plant in central Vietnam. The XGBoost model delivered the best performance with R² = 0.9998, mean absolute error (MAE) = 1.62 kWh, and root mean square error (RMSE) = 2.209 kWh, outperforming conventional methods. Additionally, the low computational demand of the developed model allows implementation on affordable hardware platforms, such as Raspberry Pi 4, enabling practical real-time monitoring and timely detection of PV performance degradation due to factors like panel defects, natural aging, and dust accumulation.