Logo image
Predicting Trfic-Induced PV Soiling Losses Using Machine Learning
Thesis   Open access

Predicting Trfic-Induced PV Soiling Losses Using Machine Learning

Jane Lubisi
M.Eng., University of Johannesburg
2025
Handle:
https://hdl.handle.net/10210/520715

Abstract

Soiling losses, caused by the accumulation of dust and particulate matter on PV module surfaces, are a major environmental factor reducing solar power generation efficiency. In South Africa, this issue is especially pronounced near unpaved gravel roads, where heavy vehicle traffic produces high concentrations of airborne dust, leading to significant revenue losses. The key operational challenge is determining an optimal, cost-effective cleaning schedule. This study evaluated the impact of gravel road proximity on PV soiling at a utility-scale installation in the North West Province. A quantitative, comparative design analysed four PV string zones at controlled distances (11 m, 39 m, 67 m, and 200 m) over an extended period (March–May 2025). Data were collected via SCADA systems for energy yield, on-site weather stations for environmental conditions, and security cameras for daily traffic intensity. These variables were used to train and validate ML models (Random Forest, XGBoost, and an Ensemble) in R Studio to predict soiling-related energy losses. Results showed that PV strings closest to the gravel road experienced the highest cumulative soiling losses, with degradation decreasing consistently with distance. The Ensemble, using the results of both models, provided the most accurate predictions, capturing the complex relationship between energy output, traffic, and distance from the dust source. This demonstrates a reproducible, data-driven methodology that enables operators to adopt dynamic, predictive cleaning strategies, focusing maintenance only on heavily soiled panels. Such an approach optimises costs, conserves water, and maximises energy yield. The study confirms the hypothesis that integrating dust source distance and traffic data with ML predictions can guide more efficient cleaning schedules, improving operational management and performance forecasting. The methodology is transferable to other PV installations in high-dust environments, supporting improved energy forecasting and day-to-day operational planning.
pdf
JL_PBv3.1 F_incl Appendix_submitted9.31 MBDownloadView
Open Access

Metrics

1 Record Views

Details

Logo image