Cross-Technology Prediction of PV Cell Output Power Using a Convolutional Hierarchical Mixture of Experts Model

Accurate prediction of photovoltaic (PV) cell power from electroluminescence (EL) images is a key enabler for automated quality assessment and performance estimation in PV manufacturing and diagnostics. However, most existing image-based deep learning models are developed and evaluated for a single PV cell technology, limiting their ability to generalize across the wide variety of cell types used in practice. This work investigates the impact of PV cell technology on power prediction accuracy and proposes a novel Convolutional Hierarchical Mixture of Experts (CHME) architecture to overcome these generalization limitations. First, convolutional neural networks and feature-based machine learning models are evaluated on multiple PV cell technologies. While technology-specific convolutional models achieve low mean absolute errors (MAEs) of 0.02–0.04 when tested on the same cell type, their performance deteriorates substantially (MAEs of 0.08–0.23) when applied to different technologies. Feature-based models exhibit greater robustness across technologies but at the cost of lower prediction accuracy. To address these limitations, the proposed CHME model combines multiple pretrained technology-specific convolutional experts with a discriminator network that automatically identifies the PV cell technology and selects the most appropriate expert for power prediction. Experimental results demonstrate that CHME achieves the best overall performance, reducing the MAE to 0.0262 compared with 0.0298 for the best standalone convolutional model, while preserving adaptability to heterogeneous datasets. These results demonstrate that explicitly accounting for PV cell technology significantly improves image-based power prediction and that the proposed hierarchical mixture-of-experts framework provides an accurate, scalable, and easily retrainable solution for real-world PV diagnostic systems.