基于生成对抗网络的后备保护在线整定快速计算方案

Quick Calculation Scheme of Backup Protection Online Setting Based on Generative Adversarial Networks

复杂环网在进行后备保护在线整定时,需花费大量的时间确定后备保护定值的配合关系,影响了在线整定的计算效率。因此,该文利用数据驱动的思想,将人工智能技术首次应用到在线整定领域,提出一种基于生成对抗网络(generative adversarial networks,GAN)的后备保护在线整定快速计算方案。首先,构建一个基于Wasserstein距离的条件GAN,以系统运行方式作为条件标签,将后备保护定值配合对转化为矩阵的索引,由此形成后备保护配合对矩阵,作为GAN训练的真实样本数据。GAN的生成器网络和判别器网络主要由卷积神经网络(convolutional neural network,CNN)组成,并对神经网络的输出进行批标准化处理,以提高训练的稳定性。然后,给出基于GAN的后备保护在线整定方案的具体实现,并采用并行计算技术加速CNN和定值的计算过程,进一步提高在线整定的计算效率。最后,以IEEE 39节点系统为例,计算各种运行方式下的后备保护配合对矩阵样本,构建GAN训练所需的数据集,对所提方案进行验证。仿真结果表明,该文所述方案通过GAN学习真实样本数据中各个后备保护定值之间复杂的配合关系,能够根据不同的运行方式给出相应的配合对矩阵,从而实现后备保护在线整定的快速计算。

The complex ring network needs much time to determine the coordination relationship of backup protection setting values during the online setting, which affects the calculation efficiency of the online setting. Therefore, this paper proposed a quick calculation scheme of backup protection online setting based on generative adversarial networks(GAN), which uses ideas of data-driven and applies the artificial intelligence technique to the online setting area for the first time. Firstly, a conditional GAN based on Wasserstein distance was constructed. It uses system operation modes as conditional labels and transforms the coordination relationship of backup protection into matrix indices to form the backup protection pair matrix as real samples for GAN training. The generator network and discriminator network of GAN were mainly composed of convolutional neural network(CNN), and their network output was batch normalized to enhance the training stability. Then, the implementation of the backup protection online setting scheme based on GAN was given, and parallel computing technology was used to accelerate the calculation process of CNN and setting values, which further improves the calculation efficiency of the online setting. Finally, a training set for GAN was built in the IEEE 39-bus system through computing backup protection pair matrix samples in varied operation modes to test the proposed scheme. The results demonstrate that the proposed scheme utilizes GAN to learn the coordination relationship of backup protection setting values in real samples, and it can give the corresponding backup protection pair matrix according to different operating modes, thus realizes the protection backup online setting quick calculation.