This work focuses on the design of modern power system controllers for automatic voltage regulators (AVR) and the applications of machine learning (ML) algorithms to correctly classify the stability of the IEEE 14 bus system. The LQG controller performs the best time domain characteristics compared to PID and LQG, while the sensor and amplifier gain is changed in a dynamic passion. After that, the IEEE 14 bus system is modeled, and contingency scenarios are simulated in the System Modelica Dymola environment. Application of the Monte Carlo principle with modified Poissons probability distribution principle is reviewed from the literature that reduces the total contingency from 1000k to 20k. The damping ratio of the contingency is then extracted, pre-processed, and fed to ML algorithms, such as logistic regression, support vector machine, decision trees, random forests, Naive Bayes, and k-nearest neighbor. A neural network (NN) of one, two, three, five, seven, and ten hidden layers with 25%, 50%, 75%, and 100% data size is considered to observe and compare the prediction time, accuracy, precision, and recall value. At lower data size, 25%, in the neural network with two-hidden layers and a single hidden layer, the accuracy becomes 95.70% and 97.38%, respectively. Increasing the hidden layer of NN beyond a second does not increase the overall score and takes a much longer prediction time; thus could be discarded for similar analysis. Moreover, when five, seven, and ten hidden layers are used, the F1 score reduces. However, in practical scenarios, where the data set contains more features and a variety of classes, higher data size is required for NN for proper training. This research will provide more insight into the damping ratio-based system stability prediction with traditional ML algorithms and neural networks.

中文翻译：

---

使用神经网络的电力系统稳定性分析

这项工作的重点是自动电压调节器 (AVR) 的现代电力系统控制器的设计和机器学习 (ML) 算法的应用，以正确分类 IEEE 14 总线系统的稳定性。与 PID 和 LQG 相比，LQG 控制器具有最佳时域特性，而传感器和放大器增益则动态变化。之后，对 IEEE 14 总线系统进行建模，并在 System Modelica Dymola 环境中模拟应急场景。从将总意外事件从 1000k 减少到 20k 的文献中回顾了蒙特卡洛原理与修正泊松概率分布原理的应用。然后提取意外事件的阻尼比，对其进行预处理，并提供给 ML 算法，例如逻辑回归、支持向量机、决策树、随机森林、朴素贝叶斯和 k 最近邻。考虑一个具有25%、50%、75%和100%数据大小的一、二、三、五、七和十个隐藏层的神经网络（NN）来观察和比较预测时间、准确性、精确度、和召回价值。在较低的数据量下，25%，在具有两个隐藏层和一个隐藏层的神经网络中，准确率分别变为 95.70% 和 97.38%。增加 NN 的隐藏层超过一秒并不会增加整体分数，并且需要更长的预测时间；因此可以丢弃进行类似分析。此外，当使用五层、七层和十层隐藏层时，F1 分数会降低。然而，在实际场景中，数据集包含更多特征和各种类别，NN 需要更大的数据量才能进行适当的训练。