A Physics-Informed Data-Driven Soiling Prediction Model for Heliostat Fields in High-Altitude CSP Plant

Abstract

Heliostat soiling significantly degrades the optical efficiency and power output of tower solar thermal power plants. Traditional cleaning strategies typically rely on fixed time intervals or reactive threshold triggers, lacking predictive foresight and adaptive scheduling capabilities under complex micro-climate conditions. To solve this issue, this paper proposes a physics-informed data-driven framework for predicting heliostat reflectance degradation, utilizing 2025 high-resolution hourly meteorological and air quality telemetry from a certain region in Qinghai, China. A physical engine considering dust deposition and dew-induced cementation is constructed to generate continuous reflectance tracking labels under strict engineering boundaries. Subsequently, an extreme gradient boosting (XGBoost) model is established to predict the 24-hour ahead reflectance based on recent physics-informed machine learning paradigms. Experimental results demonstrate that the proposed model achieves a superior Root Mean Square Error (RMSE) of 0.008176, significantly outperforming the linear regression baseline. Feature importance analysis indicates that diurnal temporal cycles and wind patterns dominate the degradation process. The proposed framework provides a robust predictive foundation for intelligent maintenance scheduling in high-altitude concentrate solar power (CSP) facilities.