A photovoltaic array including several modules in series may show mismatching due to discrepancies among the module conditions, mainly due to partial shadowing. Therefore, the shape of the current–voltage curve deeply changes with respect to the one corresponding to uniform operation. This article shows that a small set of points around the maximum power allows us to detect the occurrence of the mismatching. This approach exploits such a limited information to detect if the module is subjected to mismatching, so that the adoption of a GMPPT algorithm can be avoided. The curvature change is identified by using different machine learning techniques: decision trees, multilayer perceptrons, radial basis functions, and support vector machines. To reduce the classification error, before the fitting of the models, we implement a novel process of selection of the training samples based on a self-organizing map. This procedure makes easier the optimization of the number of hidden neurons. The support vector classifier and the multilayer perceptron with one hidden layer outperform the other approaches, being the former better than the last for extreme mismatching. However, the prediction time of this multilayer perceptron is significantly smaller than the required by the support vector machine.

中文翻译：

---

通过机器学习技术检测光伏阵列中的部分阴影和不匹配现象

包括多个串联模块的光伏阵列可能由于模块条件之间的差异而显示不匹配，主要是由于部分阴影。因此，电流-电压曲线的形状相对于均匀操作对应的形状发生了很大的变化。本文表明，最大功率附近的一小组点使我们能够检测到失配的发生。这种方法利用这种有限的信息来检测模块是否受到不匹配的影响，从而可以避免采用 GMPPT 算法。通过使用不同的机器学习技术来识别曲率变化：决策树、多层感知器、径向基函数和支持向量机。为了减少分类误差，在模型拟合之前，我们实现了一种基于自组织图选择训练样本的新过程。这个过程更容易优化隐藏神经元的数量。支持向量分类器和具有一个隐藏层的多层感知器优于其他方法，前者在极端失配方面优于后者。然而，这种多层感知器的预测时间明显小于支持向量机所需的时间。