Design of an Intelligent Framework for Dust Typing, Severity Quantification, and Performance Recovery in Photovoltaic Systems Using Multi-Domain Learning and Physics-Guided Reinforcement Control
Abstract: In dry areas, airborne dust and aerosols rapidly decrease photovoltaic (PV) performance, preventing large-scale solar deployment. Most soiling studies employ empirical soiling ratios, static threshold cleaning, or single-domain machine learning models that ignore aerosol type, adhesion, and electrical deterioration. To address these restrictions, we developed a validated multi-stage analytical framework using spectral, morphological, and operational intelligence for dust characterisation, severity assessment, removal optimization, and post-cleaning performance recovery sets. AeroSpec-Mixer, a spectral?morphology hybrid transformer, uses domain-robust co-attention from AERONET, MODIS, and SEM?EDS data to diagnose dust type in the proposed model chain. Using the latent dust-type vector, physics-guided severity estimator SAGE Index calibrates loss index sets by combining adhesion energy, optical attenuation, and IV-curve deterioration into a monotonic Gaussian process. In an adhesion-aware dynamic environment, constrained reinforcement learner CLEAN-RL schedules water- and cost-optimal cleaning. MPPT and inverter parameters are adapted using uplift-based counterfactual controller ReCap-Net to preserve post-clean gains, enhancing cleaning results. The conformal-risk evaluation module VALOR-X verifies statistically fair performance across time-split datasets and samples. The combined model improves dust-type accuracy to > 90%, loss-rate RMSE by 40%, plant production by 6% annually, and water use by half across 12 months of field and remote sensing data samples. This pipeline presents a validated analytical approach for intelligent PV soiling management and control sets.
Keywords: Dust Characterization, Photovoltaic Soiling, Reinforcement Learning, Spectral Morphology Fusion, Performance Recovery, Analysis
How to Cite?: Narendrakumar H Adkine, Dr. Sachin P. Jolhe, "Design of an Intelligent Framework for Dust Typing, Severity Quantification, and Performance Recovery in Photovoltaic Systems Using Multi-Domain Learning and Physics-Guided Reinforcement Control", Volume 14 Issue 12, December 2025, International Journal of Science and Research (IJSR), Pages: 959-974, https://www.ijsr.net/getabstract.php?paperid=SR251211080515, DOI: https://dx.doi.org/10.21275/SR251211080515