Comparison of Machine Learning Algorithms for Predicting Dust-Related Efficiency Loss in a Photovoltaic Power Plant in Semnan, Iran

Dust accumulation is one of the main environmental factors affecting photovoltaic (PV) performance, particularly in arid and semi-arid regions where long dry periods and frequent dust events can lead to noticeable efficiency degradation. This study compares machine learning algorithms for predicting dust-related efficiency loss in a photovoltaic power plant located in Semnan, Iran. A five-year daily dataset was used, including meteorological variables, dust-related indicators, PM۲.۵ concentration, solar radiation, rainfall, humidity, sunshine hours, and pvlib-based simulated daily energy output. A soiling accumulation index was first formulated to represent daily dust growth, the effects of particulate matter, dust reports and dust storm reports, and the partial natural cleaning effect of rainfall. The resulting dust-related efficiency loss index was then used as the target variable for model training and evaluation. Unlike cleaning-schedule optimization studies, the present work focuses on continuous efficiency-loss prediction rather than identifying cleaning dates or performing techno-economic assessment. Four regression algorithms, namely Random Forest, Support Vector Regression, Boosted Trees, and Artificial Neural Network, were trained using data from ۲۰۱۹ to ۲۰۲۳ and tested on unseen data from ۲۰۲۴. Model performance was evaluated using MAE, RMSE, MAPE, R², correlation coefficient, and nRMSE. Random Forest achieved the best predictive performance, with MAE = ۰.۱۳۷۶%, RMSE = ۰.۴۳۹۳%, MAPE = ۰.۵۰۳۶%, R² = ۰.۹۸۹۶, R = ۰.۹۹۴۹, and nRMSE = ۱.۴۶۴۴%. Boosted Trees also provided comparable results, while SVR and ANN showed weaker generalization during the testing period. Feature-importance analysis indicated that days since last effective rainfall, simulated daily energy output, seasonal indicators, and PM۲.۵ were among the most influential predictors. The findings suggest that tree-based ensemble models are well suited for estimating dust-related PV efficiency loss under real climatic conditions and can support data-driven performance monitoring and maintenance planning in dust-prone regions.

کلیدواژه ها:

photovoltaic systems ، machine learning ، dust accumulation ، soiling loss ، efficiency degradation ، performance prediction ، arid regions ، environmental factors