基于GAN和多通道CNN的电力系统暂态稳定评估

Transient Stability Assessment of Power System Based on GAN and Multi-channel CNN

目前基于深度机器学习的电力系统暂态稳定评估对介于稳定和失稳边界的系统状态判别存在一定困难,同时也难以兼顾在线评估的准确性和快速性。针对该问题,该文提出一种基于多通道卷积神经网络(convolutional neural network,CNN)和生成对抗网络(generative adversarial nets,GAN)的暂态稳定评估方法。首先构建了含级联多通道CNN的电力系统暂态稳定状态评估模型,通过前级多通道CNN预测非边界样本的暂态稳定状态并确定原始边界样本集;其次交替训练GAN模型的生成器和判别器以实现边界样本集增强,用增强后的边界样本集训练后级多通道CNN,使其能够可靠判别边界样本的暂态稳定状态,从而提高了状态评估的准确率;此外,在故障清除时刻预测出稳定系统的稳定程度以及失稳系统的安全控制时间裕度,从而保证了在线评估的快速性,也为后续控制策略提供一定参考。在IEEE-39节点系统和某省级电力系统上仿真表明:所提模型的评估效果相较于其他常用深度学习算法而言更为优越,在同步相量测量装置测量信息含噪声的情况下,该模型表现出较强的鲁棒性。

At present, the power system transient stability assessment based on the deep machine learning has some difficulties in distinguishing the system state between the stability and instability, which may affect the accuracy and rapidity of the on-line assessment. To solve this problem, a transient stability assessment based on the multichannel convolutional neural network(m CNN) and the generative adversarial networks(GAN) is proposed. Firstly, a power system transient stability state assessment model with the cascaded m CNN is constructed. The former m CNN is used to predict the transient stability state of non-boundary samples and determine the original boundary sample set;Secondly, the generator and discriminator of the GAN model are alternately trained to enhance the boundary sample set. The enhanced boundary sample set is used to train the later stage m CNN so that it can reliably distinguish the transient stability state of the boundary samples so as to improve the accuracy of state evaluation;In addition, the stability degree of the stable system and the safety control time margin of the unstable system are predicted at the time of fault clearing, which not only ensures the rapidity of the on-line evaluation, but also provides a certain reference for the subsequent control strategies. Simulation on IEEE-39 bus system and a provincial power system shows that the evaluation effect of the proposed model is superior to the other conventional deep learning algorithms. While there exists noise in the measurement information of the synchronous phasor measurement device, this model performs strong robustness.