基于改进生成式对抗网络的电网异常数据辨识方法

基于数据驱动的电网异常数据辨识方法已成为电网安全领域研究的重点,由于实际电力发电统计数据中异常数据样本数极少,给通过数据挖掘方法辨识异常数据情况带来了极大困难。文章提出了一种基于Wasserstein生成式对抗网络(Wasserstein generative adversarial networks, WGAN)和孤立森林算法(isolation forest,iForest)的发电统计异常数据辨识方法。首先,利用WGAN交替训练生成器和判别器学习发电统计数据的分布特性并生成样本,用生成异常样本对原始异常样本进行增强,根据异常数据辨识精度确定异常样本的扩充比例;然后,在扩充后得到的平衡数据集上利用孤立森林算法实现异常数据辨识;最后,通过扩充样本前后模型的准确率、查全率以及查准率来比较模型异常数据的辨识效果。算例结果表明,文章提出的异常样本增强方法能够有效地改善辨识模型对于多数类的分类偏好问题,提升整体辨识精度。

An Abnormal Data Identification Method Based on Improved Generative Adversarial Network

The identification method based on data driving for identifying abnormal data in power grid has become the focus of research in the field of power grid security. Due to the small number of abnormal data in the actual statistical data of power generation, it is extremely difficult to identify abnormal data through data mining. This paper proposes an improved generative adversarial network and isolated forests based abnormal data identification method. Firstly, by using WGAN alternating training generator and discriminator, the distribution characteristics of power generation statistical data are learned and samples are generated, which will generate abnormal samples to enhance the original abnormal samples. According to the accuracy of outlier data identification, the expansion ratio of outlier samples is determined. Then the isolated forest algorithm is used to identify the abnormal data on the expanded balanced dataset. Finally, an evaluation index system consisting of accuracy, recall and precision is selected to evaluate and compare the identification effects of models before and after category equalization. The results show that the proposed method can effectively improve the classification preference of the identification model for most classes and improve the overall identification accuracy.