Over the past few years, with the access of large-scale new energy sources, the problem of subsynchronous oscillation (SSO) in power systems has presented a novel multisource and multitransformation form, which may be significantly threatening. Conventional control and protection methods primarily give rise to device protection actions in the presence of severe oscillation. On the whole, online monitoring only identifies the frequency and amplitude, whereas it cannot identify the attenuation factor. Moreover, the determination of the warning threshold is more dependent on human experience, so the reliability and rapidity of the early warning cannot be ensured. This study conducts an in-depth investigation of the wind-thermal power bundling and extreme high-voltage alternating current- (AC-) direct current (DC) hybrid transmission system. The major factors of SSO using this system are unclear, which brings difficulties to effective monitoring. Given the mentioned problems, a method combining Levenberg–Marquardt- (LM-) Backpropagation (BP) machine learning and Sensitivity Analysis (SA) and principal component analysis (PCA) is developed. First, the sensitivity analysis of each factor in the system is conducted to identify the major factors of SSO. Subsequently, the historical sample data are reduced with the principal component analysis to reduce the redundancy, which is adopted to train the regression model to determine the attenuation factor and frequency and then send them to the classifier for classification to complete the task of the assessment model. When a novel data signal is uploaded, the assessment model identifies the attenuation factor and frequency and subsequently determines the presence of SSO. Accordingly, an early warning is conducted. The system's refined simulation model and machine learning model verify the effectiveness of the method.

中文翻译：

---

Levenberg-Marquardt和反向传播算法结合灵敏度分析和主成分分析的次同步振荡在线监测和预警

在过去的几年中，随着大规模新能源的使用，电力系统中的次同步振荡（SSO）问题提出了一种新颖的多源多变换形式，这可能会带来严重威胁。常规的控制和保护方法主要是在严重振荡的情况下引起设备保护动作。总体而言，在线监控只能识别频率和幅度，而不能识别衰减因子。而且，警告阈值的确定更多地取决于人类的经验，因此不能确保预警的可靠性和快速性。这项研究对风热发电捆绑和特高压交流电（AC-）直流电（DC）混合传输系统进行了深入研究。使用该系统的SSO的主要因素尚不清楚，这给有效监控带来了困难。针对上述问题，开发了一种结合Levenberg-Marquardt-（LM-）反向传播（BP）机器学习，灵敏度分析（SA）和主成分分析（PCA）的方法。首先，对系统中每个因素进行敏感性分析，以确定SSO的主要因素。随后，通过主成分分析对历史样本数据进行约简，以减少冗余，然后采用该模型训练回归模型确定衰减因子和频率，然后将其发送到分类器进行分类，以完成评估模型的任务。 。上载新的数据信号时，评估模型确定衰减因子和频率，然后确定SSO的存在。因此，进行预警。系统改进的仿真模型和机器学习模型验证了该方法的有效性。