A MULTI-HYBRID DEEP LEARNING FRAMEWORK INTEGRATING TRANSFER LEARNING AND ENSEMBLE CLASSIFICATION FOR AUTOMATED PHOTOVOLTAIC PANEL DEFECT RECOGNITION

Abstract

The rapid expansion of photovoltaic (PV) installations has increased the need for intelligent and automated defect detection systems capable of maintaining energy efficiency and operational reliability. Manual inspection of PV panels is time-consuming, labor-intensive, and prone to inaccuracies, particularly in large-scale solar farms. This study proposes a Multi-Hybrid Deep Learning Framework integrating transfer learning, deep feature extraction, and ensemble machine learning techniques for automated PV defect classification. Four models, namely PV-DenseFineNet, PV-DenseXGBHybridNet, PV-EfficientXGBHybridNet, and PV-MobileRFHybridNet, were developed using DenseNet121, EfficientNetV2B0, and MobileNetV3Large backbones combined with Softmax, XGBoost, and Random Forest classifiers. Experiments were conducted on a publicly available PV panel defect dataset containing six defect categories: Bird-drop, Clean, Dusty, Electrical-damage, Physical-damage, and Snow-covered. Comparative evaluation using Accuracy, Precision, Recall, F1-Score, ROC-AUC, and confusion matrix analysis demonstrated that PV-DenseFineNet achieved superior performance with 89.5% test accuracy and 0.987 ROC-AUC. The results confirm the effectiveness of end-to-end transfer learning for intelligent photovoltaic defect recognition and solar panel monitoring applications.