随着直流输电技术的快速发展和大规模应用，电力系统的故障特征和暂态稳定模式变得更加复杂，交直流系统高度耦合、相互影响，放大了故障对电力系统运行的冲击范围和强度，给电力系统故障评估和安全运行带来巨大挑战。本文针对大规模交直流系统暂态稳定特性，提出了一种基于深度神经网络（deep neural network，DNN）的交直流系统暂态稳定评估方法。首先，提出一组反映交直流系统暂态稳定特性的121维特征集作为DNN模型的输入，构建含4层隐含层的6层DNN模型，采用Adam算法提高DNN的收敛速度和运行效率，以稳定和不稳定为输出结果，利用深度学习架构对特征数据集与稳定结果之间的映射关系进行训练。最后，利用改进的IEEE新英格兰10机39节点作为测试系统对所提方法进行验证，结果表明所提出的基于DNN的交直流暂态稳定评估方法准确率可达98%，比基于支持向量机、决策树等的暂态稳定评估方法准确率和收敛性更好，验证了本文所提方法的正确性和实用性。

With the rapid development and large-scale application of HVDC transmission technology, the fault characteristics and transient stability modes of power systems become more complex. Hybrid AC/DC power systems are highly coupled and interact with each other, amplifying the impact range and intensity of faults on power system operation, which poses great challenges to stability assessment and safe operation of hybrid AC/DC systems. Therefore, to deal with these challenges, this paper proposes a transient stability evaluation method for large-scale AC/DC systems based on the deep neural network(DNN). Firstly, a set of 121 characteristic values reflecting transient stability behaviours of AC/DC systems is proposed and a 6-layer DNN model including 4-layer hidden layer is constructed. The Adam algorithm is used to improve the convergence speed and operation efficiency of the DNN. Stable and unstable are considered as the output results, and the deep learning architecture is used to train the mapping relationship between feature data sets and stable results. Finally, the proposed method is tested by using the improved IEEE New England 10 generator 39 bus system. Results show that the accuracy of the proposed method is up to 98%, which is higher than that of the transient stability evaluation methods based on support vector machine (SVM), decision tree (DT), etc., so as to verify the feasibility and practicability of the proposed method.

 摘要：在机器学习领域，暂态稳定评估问题被定义为通过大量故障样本来估计稳定边界的二分类问题。该文提出了一种深度学习方法来解决这个二分类问题。该方法包含4个步骤：首先，利用样本数据构建原始输入特征来描述电力系统动态特性；然后，采用变分自动编码器（variational auto-encoders,VAE）对原始输入特征进行无监督学习实现特征抽取，从而获得高阶特征；之后，对卷积神经网络（convolution neural network,CNN）进行有监督学习训练得到高阶特征与电力系统暂态稳定性之间的映射关系；最后，将训练得到的模型应用于电力系统在线暂态稳定评估。在新英格兰39节点测试电力系统的仿真试验表明，所提出的暂态稳定评估（transient stability assessment，TSA）模型具有评估精度高、不稳定样本评估错误率低、抗噪声干扰能力强的特点，适合基于广域测量信息的准实时在线暂态稳定评估。  

&nbsp;ABSTRACT： In the field of machine learning, transient stability assessment can be considered as a two-class problem of estimating the stability boundary through large number of fault samples. This paper proposes a method of deep learning to solve this problem. The method consists of four stages: firstly, using samples to construct the original input feature for describing the dynamic characteristics of the power system；secondly, variational auto-encoders （VAE） is used to perform unsupervised learning on the original input feature to obtain high-order features；thirdly, the supervised training of convolution neural network （CNN） is carried out to obtain the relationship between high order characteristic and transient stability of power system；finally, the model is applied to the transient stability assessment of power system. Simulation on the New England 39-bus test system shows that the proposed approach has high accuracy, rare misclassification of unstable sample and excellent robustness with noise for transient stability assessment （TSA）. Therefore, it is suitable for quasi-real-time online transient stability assessment based on wide-area measurement information.<div>&nbsp;</div>

After a very fast and efficient discriminative broad learning system (BLS) that takes advantage of flatted structure and incremental learning has been developed, here, a mathematical proof of the universal approximation property of BLS is provided. In addition, the framework of several BLS variants with their mathematical modeling is given. The variations include cascade, recurrent, and broad-deep combination structures. From the experimental results, the BLS and its variations outperform several exist learning algorithms on regression performance over function approximation, time series prediction, and face recognition databases. In addition, experiments on the extremely challenging data set, such as MS-Celeb-1M, are given. Compared with other convolutional networks, the effectiveness and efficiency of the variants of BLS are demonstrated.

在暂态功角稳定评估和暂态电压稳定评估的相关研究中,通常分别构建独立的评估模型,这阻碍了不同任务间的信息共享,浪费了计算和存储资源。考虑不同评估任务间往往存在相似性和差异性,为更好地实现二者同时评估,文章提出了一种基于混合门控循环单元(gated recurrent unit, GRU)子层的多任务暂态稳定评估模型。由于电力系统暂态过程具有明显的时序特性,模型采用GRU子层高效地提取量测数据中的时序特征;并在模型结构中引入门控机制以自动调节各个子层在构建不同评估任务特征表示时的占比,不仅促进了不同任务间的信息共享,还削弱了不同任务间差异性对模型训练的负面影响。在IEEE 39节点测试系统中的仿真实验结果表明,文章提出的模型具更好的评估性能和计算速度。

In the related research of transient power-angle stability assessment and transient voltage stability assessment, independent assessment models are usually constructed, respectively, which hinders the information sharing among different tasks and wastes computing and storage resources. Considering the similarities and differences among different assessment tasks, in order to better realize the two assessments at the same time, a multi-task transient stability assessment model based on sub-layer of hybrid gated recurrent unit (GRU) is proposed in this paper. Because the power system transient process has obvious time-series characteristics, the GRU sub-layer is used to extract the time-series characteristics of the measured data efficiently. The gating mechanism is introduced into the model structure to automatically adjust the proportion of each sub-layer in constructing the feature representation of different evaluation tasks, which not only promotes the information sharing among different tasks, but also weakens the negative impact of the differences among different tasks on model training. The experimental results in IEEE 39-bus test system show that the model proposed in this paper has better evaluation performance and computing speed.

By focusing on the accuracy limitations of the naive Bayesian classifier in the transient stability assessment of power systems, a tree augmented naive Bayesian (TAN) classifier is adopted for the power system transient stability assessment. The adaptive Boosting (AdaBoost) algorithm is used in the TAN classifier to form an AdaBoost-based tree augmented naive Bayesian (ATAN) classifier for further classification performance improvement. To construct the ATAN classifier, eight attributes that reasonably reflect the transient stability or transient instability of a power system are selected as inputs of the proposed classifier. In addition, the class-attribute interdependence maximization (CAIM) algorithm is used to discretize the attributes. Then, the operating mechanism of the power system is used to obtain the dependencies between the attributes, and the parameters of the ATAN classifier are learned according to the Bayes’ theorem and the criterion of maximizing a posterior estimation. Four evaluation indicators of the ATAN classifier are used, that is, the value of Kappa, the area under the receiver operating characteristic curve (AUC), F1 score, and the average evaluation indicator. Lastly, experiments are implemented on the IEEE 3-generator 9-bus system and IEEE 10-generator 39-bus system. The simulation results show that the ATAN classifier can significantly improve the classification performance of the transient stability assessment of the power system.