Dataset of "A Physics-guided, Environmentally Compensated, Intelligent and Noise-Resilient Framework to Multi-Class Fault Detection and Diagnosis in Photovoltaic Systems Including Shunt and Series Resistances Degradation 核心 · 已核验

mxds10e08k43sd42

In photovoltaic (PV) systems, intrinsic fault signatures are often masked by measurements and environmental variability, making reliable fault detection and diagnosis particularly challenging. To address this issue, this paper proposes a physics-guided Machine Learning framework that combines PV system modeling and data-driven intelligence to enhance classification performance. In order to replicate realistic behavior, reproduce degradation-related faults, and generate a physically consistent dataset of defective modules using the equivalent-circuit parameter changes, specifically the series resistance (Rs) and shunt resistance (Rsh), a complete PV model using Simulink was developed. Building on this foundation, Environmentally Compensated Features were incorporated to enhance class separability and mitigate sensitivity to environmental variations.

A three-stage, physics-guided framework for PV fault detection and diagnosis is proposed. First, synthetic data are generated using physical models. Second, faults are detected using an AdaBoost ensemble, with hyperparameters tuned via Bayesian optimization for binary classification. Third, faults are diagnosed using a nonlinear Support Vector Machine (SVM), which is also optimized via Bayesian search, while LASSO (Least Absolute Shrinkage and Selection Operator)- based feature selection reduces dimensionality for the multi-classification task. To ensure unbiased performance assessment and prevent data leakage, models are evaluated using nested cross-validation. Results indicate improved separability of fault signatures and reliable classification of six fault categories, including degradation-related faults such as Rs and Rsh deviations that are typically difficult for purely data-driven methods. The detection model attains 99.74% accuracy, 99.78% balanced accuracy, and an F1 score of 0.997. The diagnosis model achieves 99.68% accuracy and an F1 score of 0.9967 across six classes. A robustness analysis under additive Gau

本卡描述如含来自官方页的原文片段,其版权归原作者,不在本站 CC-BY 4.0 许可范围(见关于与许可)。

落地页

https://zenodo.org/doi/10.5281/zenodo.20407863

许可证

CC-BY-4.0 (置信:verified_official)

国内可访问性

国内直连:直连超时(慢或不稳定,非封锁证据) (2026-07-11 检测) 非直连:可达 (2026-07-11 检测)
「直连超时」表示检测窗口内未完成,系慢或不稳定证据,不构成封锁证据。

发布年份

2025

发布方

Zenodo

PHM 任务

fault_detection fault_diagnosis

运行工况