Advanced analysis and monitoring of photovoltaic solar modules is required to maintain the reliable operations of photovoltaic plants. Hence, it requires diagnostics through current–voltage (IV) curves, electroluminescence (EL) imaging, and other measurement techniques. The analysis through IV characterization provides the discerning insight about the quantitative measure of solar module performance, while the image characterization methods on EL images can capture spatial defects with microscopic resolution such as microcracks, broken cells interconnections, shunts, among many other defect types. The fusion of these two methods with supervised and unsupervised machine learning can generate unique insight with classification, regression, and dimension reductions on IV–EL data. In this study, we have performed the IV–EL correlation by classifying the IV data based on EL image annotation (where the class information is coming from EL image). The feature vectors consist of IV curve parameters and statistical features. We have first applied the unsupervised learning algorithms t-distributed stochastic neighbor embedding (t-SNE) and uniform manifold approximation and projection (UMAP) for dimensionality reduction to understand the importance of various features on EL defect types. Furthermore, we had applied feature selection algorithms before applying the classification algorithms. We have performed the classification of various defect types by applying the random forests (RF) and XGBoost algorithm while identifying the top features. The accuracy was achieved greater than 91% and 95%, respectively, for supervised methods on the top five features. This correlation of IV–EL measurement could benefit in quick identification of various defect types in PV modules with only IV curve parameters, given the classification models are modeled using large-scale datasets and tuned optimally.

中文翻译：

---

使用机器学习算法对光伏模块的 IV 曲线和电致发光图像进行高级分析

需要对光伏太阳能组件进行高级分析和监控，以维持光伏电站的可靠运行。因此，它需要通过电流-电压 (IV) 曲线、电致发光 (EL) 成像和其他测量技术进行诊断。通过 IV 表征的分析提供了关于太阳能模块性能定量测量的敏锐洞察力，而 EL 图像上的图像表征方法可以捕捉具有微观分辨率的空间缺陷，例如微裂纹、破损的电池互连、分流器以及许多其他缺陷类型。这两种方法与有监督和无监督机器学习的融合可以通过 IV-EL 数据的分类、回归和降维产生独特的见解。在这项研究中，我们通过基于 EL 图像注释（其中类信息来自 EL 图像）对 IV 数据进行分类来执行 IV-EL 相关性。特征向量由 IV 曲线参数和统计特征组成。我们首先应用了无监督学习算法t -分布随机邻域嵌入 ( t-SNE) 和统一流形逼近和投影 (UMAP) 用于降维以了解各种特征对 EL 缺陷类型的重要性。此外，我们在应用分类算法之前应用了特征选择算法。我们通过应用随机森林 (RF) 和 XGBoost 算法对各种缺陷类型进行了分类，同时识别了最重要的特征。对于前五个特征的监督方法，准确率分别超过 91% 和 95%。鉴于分类模型是使用大规模数据集建模并优化调整的，IV-EL 测量的这种相关性有助于快速识别仅具有 IV 曲线参数的光伏模块中的各种缺陷类型。