Abstract Loss of massive generation caused by HVDC blocking or tripping of large power plants is a severe threat to the frequency security of receiving-end grids, especially those with a high penetration level of renewable generation. Due to the uncertainty of renewable generation, a large number of scenarios need to be checked for hour-ahead frequency security assessment, which is time-consuming if they are assessed with full time-domain simulation. To improve assessment efficiency, a maximum frequency deviation assessment method with clustering based on Metric Learning (ML) is proposed in this paper. The distance measure of samples is firstly adjusted by ML with Kernel Regression to make samples with similar frequency dynamics close to each other. The training process of ML is optimized with Mini-Batch Gradient Descent (MBGD) method. Then fuzzy k-means clustering is used to group the samples into multi clusters by maximizing the membership degree of each training sample with the distance measure adjusted by ML. Finally, corresponding to each cluster, a Support Vector Regression (SVR) model is established to build the relationship between maximum frequency deviation and steady-state power flow features, and Core Vector Regression (CVR) is used to accelerate the training process of SVR. The membership degree of the new scenario to assess with respect to each cluster is calculated according to the distance measure learned by ML to classify it into a specific cluster in the on-line assessment process. The corresponding SVR model is used to assess the maximum frequency deviation of the new scenario. The New England 39-bus system and a simplified provincial power system of China are adopted to verify the validity of the proposed assessment method.

中文翻译：

---

基于度量学习的聚类最大频偏评估

摘要 大型发电厂直流阻塞或跳闸造成的大量发电损失严重威胁受端电网的频率安全，尤其是可再生能源发电普及率较高的受端电网。由于可再生能源发电的不确定性，需要对大量场景进行提前一小时的频率安全评估，如果采用全时域模拟进行评估，则非常耗时。为了提高评估效率，本文提出了一种基于度量学习（ML）的聚类最大频偏评估方法。样本的距离度量首先通过 ML 和核回归进行调整，使具有相似频率动态的样本彼此接近。ML 的训练过程使用 Mini-Batch Gradient Descent (MBGD) 方法进行了优化。然后使用模糊k-means聚类，通过使用ML调整的距离度量，最大化每个训练样本的隶属度，将样本分组为多个簇。最后，对应每个簇，建立支持向量回归（SVR）模型建立最大频偏与稳态潮流特征的关系，并使用核心向量回归（CVR）加速SVR的训练过程。在线评估过程中，根据ML学习到的距离度量计算新场景对每个集群进行评估的隶属度，将其分类到特定的集群中。相应的 SVR 模型用于评估新场景的最大频率偏差。