Condition Monitoring of Photovoltaic Panels Through Electrical Impedance Spectroscopy and Machine Learning

Reliable operation of photovoltaic (PV) assets demands early detection of subtle faults and environmental stressors that degrade energy yield. This work investigates a data-driven condition-monitoring pipeline that fuses Electrical Impedance Spectroscopy (EIS) with modern machine learning to classify PV module temperature classes and degradation types (including dust deposition and microcracks). EIS measurements are transformed into informative descriptors such as frequency-domain magnitude/phase trends, Nyquist/Bode curve features, and equivalent-circuit parameters and fed to a diverse model suite. We evaluate TabNet (for end-to-end representation learning on tabular spectra), AutoML frameworks (H2O AutoML, AutoGluon) for systematic model search, and an ensemble StackingClassifier combining LightGBM, SVM, and CatBoost for robust decision-level fusion. To enable trustworthy deployment, we apply SHAP and LIME to interpret class-driving spectral regions and component parameters. For edge use cases, we further develop a lightweight neural network tailored for embedded inference on resource-constrained devices. Across varied operating conditions, the proposed approach achieves accurate and stable classification while maintaining low latency and model size suitable for on-site monitoring. The results highlight the viability of EIS-centric, explainable ML for proactive PV asset management, enabling predictive maintenance scheduling and reduced lifecycle costs.

Keywords:

Photovoltaic Panels, Electrical Impedance Spectroscopy (EIS), Condition Monitoring, Temperature Classification, Degradation Detection, Dust, Microcracks, TabNet, LightGBM, Support Vector Machine (SVM), CatBoost, StackingClassifier, Explainable AI, SHAP, LIME, Embedded Inference, Edge Computing, Nyquist/Bode Features, Equivalent Circuit Modeling, Predictive Maintenance.