Hybrid spectral learning framework for real-time material classification of end-of-life photovoltaic panels

The increasing deployment of photovoltaic (PV) systems has resulted in a significant rise in end-of-life (EOL) solar panels, creating major challenges in material identification, recycling, and sustainable waste management. This work proposes a hybrid digital twin–guided spectral learning framework for intelligent classification of photovoltaic panel materials. The proposed framework integrates photovoltaic system simulation, spectral feature analysis, graph-based Bayesian learning, and African Vultures Optimization Algorithm (AVOA)-based parameter optimization for accurate material classification. A photovoltaic system consisting of a PV array, perturb-and-observe (P&O) maximum power point tracking controller, pulse width modulation unit, and DC–DC boost converter is modelled in MATLAB/Simulink to generate electrical and environmental signals. The generated signals are used to construct a spectral dataset containing 3258 samples with 15 features. In addition, real-time photovoltaic voltage, current, power, temperature, and solar irradiance measurements are collected from a rooftop PV system to improve practical relevance. Since hyperspectral sensing infrastructure is unavailable in the hardware setup, spectral features are obtained from simulation while electrical and environmental parameters are acquired from real-time measurements, thereby forming a hybrid dataset. The proposed digital twin spectral graph kernel Bayesian framework effectively captures nonlinear relationships among photovoltaic material features, while AVOA improves parameter optimization and convergence performance. Experimental evaluation demonstrates that the proposed model achieves 96.84% classification accuracy, 0.967 precision, 0.958 recall, and 0.962 F1-score, outperforming Random Forest, Decision Tree, Support Vector Machine, and Artificial Neural Network classifiers. The AVOA optimization achieved faster convergence within 68 iterations compared with PSO (95 iterations) and GA (120 iterations). Furthermore, the proposed framework achieved 92.15% classification accuracy under real-time operating conditions, demonstrating reliable and robust photovoltaic material classification performance. The results confirm that the proposed hybrid digital twin framework provides an efficient, scalable, and practical solution for intelligent photovoltaic material classification and sustainable end-of-life solar panel recycling.