A lightweight multi-scale deep learning approach for robust photovoltaic fault detection under noisy conditions
Year: 2026
Type: preprint
Abstract: Reliable fault diagnosis in photovoltaic (PV) systems under noisy and high-dimensional operating conditions remains challenging pattern-recognition problem. This study proposes BLGSNet, a lightweight hybrid architecture that integrates parallel multi-scale one-dimensional convolutional branches with a bidirectional long short-term memory (BiLSTM) network to capture local fault patterns and bidirectional contextual relationships across ordered PV features. BLGSNet is trained using all 31 standardized input features, while MLP-based permutation ranking and principal component analysis (PCA) are employed to analyze feature relevance and construct compact representations for conventional machine-learning comparisons. Robustness is evaluated under compositeGaussian, bias, and impulsive perturbations. On a six-class PV fault dataset, BLGSNet achieves 99.36% accuracy under noise-free conditions and maintains accuracies of 98.08% and 94.87% at 10% and 15% noise, respectively. Evaluation on an additional experimental dataset with rule-based pseudo-labels yields 98.26% accuracy, demonstrating consistent performance under practical operating variability. A laboratory-scale implementation using an NVIDIA Jetson Nano and a real-time dashboard is also developed for fault-free, string-fault, partial-shading, and component-related operating conditions. A rule-based maintenance decision-support module translates model predictions into severity levels, maintenance priorities, and recommended inspection actions. The results demonstrate the potential of BLGSNet for computationally efficient and noise-resilient PV fault monitoring.
Source: Research Square
Institution: Kookmin University
Language: English
Field: Energy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204132229