A CFD-based spatiotemporal mapping of dust deposition on solar fields using unsupervised clustering for targeted cleaning

Conventional mirror cleaning strategies for solar power plants are water-intensive and costly, so targeted cleaning is desired by identifying the dust-concentrated solar collectors. This paper utilizes a two-phase computational fluid dynamics (CFD) approach integrated with unsupervised clustering to predict the dust distribution on the solar field at a 100 MW-scale concentrated solar power (CSP) plant. The meteorological data (wind speed, direction, and relative humidity) recorded over a full year and geometrical parameters (size, layout, and number of solar collectors) are sourced directly from an operational CSP plant to develop a high-fidelity simulation model. Two distinct cleaning cycles (4 days each) for solar collectors with and without sandstorm are evaluated for model validation. Our results indicate that the root mean square error (RMSE) of model for the regular cleaning cycle varies from 7.1% to 8.1%, while for the irregular cleaning cycle (with sandstorm), it varies from 7.2% to 12.4% between the first and the last day of the cleaning cycle respectively. The proposed CFD approach is employed to generate spatially varying reflectivity data across weather conditions (varying wind speeds, humidity levels, and wind directions), effectively eliminating the need of expensive experimental setups. Unsupervised clustering is then utilized to classify the inherent trends between the meteorological conditions and reflectance loss. Based on this analysis, a dust distribution model featuring five distinct classes is developed to predict dust accumulation on solar mirrors throughout the year, enabling sustainable targeted cleaning.