Predictive modeling of photovoltaic system cleaning schedules using machine learning techniques

Abstract

Photovoltaic (PV) solar systems are a key contributor to sustainable energy generation, but their performance is significantly reduced by dust accumulation, highlighting the need for proper cleaning. This study develops predictive models to optimize cleaning schedules by forecasting the Performance Ratio (PR), a standardized metric essential to performance-guaranteed contracts. The first model uses time-series approaches (LSTM, ARIMA, SARIMAX) to predict PR, while the second uses a threshold-based ensemble voting classifier (RF, Logistic Regression, GBM) to predict cleaning needs. Two large datasets from case studies in the UAE and Jordan were used for validation. Results show SARIMAX outperforming other models, with R² values of 93.36 % and 91.74 %. The cleaning classification model achieved accuracies of 91 % and 88 % in the respective case studies. The PR prediction models outperformed the cleaning classification models in terms of accuracy. The study also identified location-specific factors influencing PV system performance, emphasizing the need for geographically tailored maintenance strategies. This research provides valuable insights for improving the efficiency and sustainability of PV systems.