Predicting PV Module Thermomechanical Durability (TC600) from BOM Features Using a TabNet-Stacked Ensemble Approach

Thermomechanical degradation is a major reliability concern in photovoltaic (PV) modules, with the TC600 test serving as a critical but time- and resource-intensive qualification standard. This study presents a TabNet-stacked ensemble learning framework for predicting TC600 power loss directly from Bill-of-Materials (BOM) features, offering a fast and interpretable alternative to physical testing. The ensemble combines four base models (Random Forest, Extra Trees, Gradient Boosting, XGBoost) with a TabNet meta-learner that provides built-in interpretability through attention mechanisms. Trained on 240 PV module designs with 22 BOM features, the model achieves excellent predictive performance (RMSE = 0.240%, MAE = 0.178%, R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.980). Feature importance analysis highlights the Cell domain as most influential (23.6%), followed by Metallization (19.3%), Glass Layer (17.3%), and Encapsulant (13.2%). The meta-learner confirms balanced contributions from all base learners. Overall, this framework enables rapid, interpretable BOM optimization for thermomechanical reliability and establishes a basis for digital PV qualification frameworks applicable to other cell types, degradation and testing protocols.