Artificial Intelligence-Based Ensemble Learning and Environmental Data Fusion for Dust-Driven Soiling-Risk Classification in Photovoltaic Systems

Soiling, or dust accumulation, is a significant challenge affecting the performance and reliability of photovoltaic (PV) systems in arid regions. This paper introduces Fusion for Understanding Soiling Events (FUSE), a data-driven framework for classifying atmospheric dust-loading conditions associated with potential PV soiling risk. The framework employs a dual-level fusion architecture. The first level integrates environmental information from globally available datasets, including solar radiation, meteorological variables, and atmospheric particulate matter, for two Egyptian locations, Cairo and Aswan. The second level combines machine learning classifiers through weighted soft voting, with weights determined by cross-validation F1-scores. The framework was trained on observations from 2019 to 2022 and evaluated on an independent 2023 temporal holdout from the same two locations. For same-day environmental risk assessment, FUSE achieved a weighted F1-score of 91.65% and a ROC-AUC of 0.9324, outperforming Random Forest in weighted F1 while showing comparable performance to XGBoost and LightGBM. Wind and particulate matter variables were among the most informative predictor groups, while solar variables contributed relatively little. At the default classification threshold, the ensemble identified 53% of the higher-dust-loading proxy events, increasing to 68% using a decision threshold selected from training data. The primary framework provides retrospective same-day assessments after the required environmental observations become available. A separate next-day prediction experiment using preceding-day environmental features achieved a weighted F1-score of up to 90.77% on the 2023 test set. The resulting classifications represent AOD-derived atmospheric dust-loading proxy labels rather than independently measured PV module soiling. Although the framework may be adaptable to other regions with comparable data availability, geographic transferability beyond Cairo and Aswan has not been evaluated. These findings suggest that environmental data fusion may contribute to future condition-based PV maintenance and Prognostics and Health Management applications, subject to independent module-soiling validation and techno-economic assessment.