Shading is one of the most common anomalies in photovoltaic (PV) systems, leading to power loss and hotspot phenomenon. Currently, most works can only realize shading detection but cannot further diagnose the type and severity of shading. This paper proposes an effective method for diagnosing shading types combining I-V curve imaging with two-stream deep neural networks (DNN), and estimating severity of five common types of shading in actual operating PV systems. In this method, the I-V curves of PV strings are first resampled and converted to standard test conditions (STC) for eliminating the effects of data scale and environmental factors on shading diagnosis results. Then, a time series imaging method called Gramian angular summation field (GASF) is used to enhance the features of shading. Additionally, a two-stream DNN combining long short-term memory (LSTM) and improved two-dimensional convolutional neural network (2D-CNN) is developed to integrate the characteristic information of I-V curves and 2D images. Furthermore, combining the PV mechanism models and characteristics of I-V curves, this work further estimates the severity of different types of shading in operating PV systems considering the effects of aging loss. The effectiveness and generalization of the proposed method are validated via simulated and experimental data obtained from simulation model and an actual PV platform.

中文翻译：

---

通过 I-V 曲线成像和双流深度神经网络诊断光伏系统中的阴影类型和严重程度


阴影是光伏 （PV） 系统中最常见的异常现象之一，会导致功率损失和热点现象。目前，大多数作品只能实现阴影检测，无法进一步诊断阴影的类型和严重性。本文提出了一种结合 I-V 曲线成像和双流深度神经网络 （DNN） 来诊断阴影类型的有效方法，并估计实际运行光伏系统中五种常见阴影类型的严重程度。该方法首先对光伏组串的 I-V 曲线进行重采样，并将其转换为标准测试条件 （STC），以消除数据尺度和环境因素对阴影诊断结果的影响。然后，使用一种称为 Gramian 角度求和场 （GASF） 的时间序列成像方法来增强阴影的特征。此外，开发了一种结合了长短期记忆 （LSTM） 和改进的二维卷积神经网络 （2D-CNN） 的双流 DNN，以整合 I-V 曲线和 2D 图像的特征信息。此外，结合 PV 机制模型和 I-V 曲线的特性，这项工作进一步估计了考虑到老化损失的影响，在运行 PV 系统中不同类型阴影的严重程度。通过仿真模型和实际光伏平台获得的仿真和实验数据验证了所提方法的有效性和泛化性。