基于双向长短时记忆网络和卷曲神经网络的电力系统暂态稳定评估

基于机器学习方法的暂态稳定评估已成为电力系统分析与控制领域的热点,由于实际系统中存在不能实现PMU的全面覆盖以及数据采集存在噪声的问题,使得传统机器学习方法的评估性能受到较大限制。针对此,构建了一种在PMU最优布点上的时间序列特征,提出了一种将改进卷积神经网络(ICNN)与双向长短时记忆网络(BiLSTM)进行融合的评估方法。该方法首先利用BiLSTM提取电压、相角以及有功功率三种基本电气量的时间序列特征,随后通过卷积和池化操作对数据进行进一步的数据挖掘,最后利用轻量梯度提升机完成对数据的分类。为了避免出现过拟合现象,该方法还通过正则化、Dropout等方式提升模型的泛化性能。在新英格兰10机39节点上的算例表明,该方法能利用基本电气量数据进行暂态稳定评估,且在复杂条件下仍能保持较好的评估性能。

Power System Transient Stability Assessment Based on BiLSTM-ICNN

Transient stability assessment based on machine learning has become the focus of research in the field of power system analysis and control. However, due to the problems of inadequate PMU coverage and noisy data acquisition, the performance of traditional machine learning methods is greatly limited. To solve this problem, this paper constructed a time series feature based on the optimal PMU placement, and presented a TSA method using a hybrid bidirectional LSTM and improved CNN architecture by the first time. The method firstly use BiLSTM to extract the time series features of voltage, phase angle and active power, and then convolution and pooling operations are used to further mine the features of the data. Finally, using the LGBM to classify the data. In order to avoid over-fitting, regularization and Dropout algorithm are used to improve the generalization performance of the model. The simulation results on New England 39-bus system showed that the proposed method could maintain better performance in the complex conditions by use of basic power system data.