Hadoop架构下基于分布式粒子群算法的暂态稳定评估特征量选择

Feature Selection for Transient Stability Assessment Applying Distributed PSO Algorithm Based on Hadoop Architecture

特征量选择是基于机器学习的电力系统暂态稳定评估的重要环节。针对现有特征量选择方法存在分类判据选择效果不佳和初始特征集构建不全面等问题,提出一种基于改进分类判据和考虑单机特征的特征量选择方法。首先以基于类内类间离散度的分类判据为基础,对类内类间离散度进行改进,同时基于信息熵提出特征熵的概念用于衡量低维特征组合中各特征量在初始特征集中的重要程度,进一步提出基于改进类内类间离散度和特征熵的分类判据;其次,利用系统特征和可表征临界机组特性的单机特征构建初始特征集,且为尽量避免所提特征量选择方法出现维数灾问题,提出用于特征量选择的Hadoop架构下分布式粒子群算法;最后,以EPRI-36节点系统和某实际系统为算例验证所提方法的有效性。

Feature selection is an important part of power system transient stability assessment based on machine learning.Aiming at the problems of existing methods of feature selection,such as poor selection effect of classification criteria and incomprehensive construction of initial feature set,a feature selection method based on improved classification criterion and single machine features is proposed.Firstly,.on basis of the classification criteria based on within-class and between-class scatter,the within-class and between-class scatter matrix are improved.Meanwhile,the feature entropy based on information entropy is proposed to measure importance of each feature of low dimensional feature combination in initial feature set,then the classification criterion based on the improved within-class and between-class scatter is proposed.Secondly,the system feature and single feature characterizing supercritical units are used to form the initial feature set.Moreover,in order to avoid dimensionality curse of the proposed feature selection method as far as possible,a distributed particle swarm optimization algorithm based on Hadoop architecture is proposed for feature selection. Finally,the proposed method is examined on the data of EPRI 36system and an actual system to demonstrate its efficiency.