基于压缩感知的电能质量扰动稀疏表示多分类研究

The time-frequency methods such as wavelet transform and s transform have been proposed in recent past, for detecting and extracting features of various categories of power quality disturbances (PQDs). However, time-frequency methods are incompetent to fulfill rapid detection although fast algorithms due to the high computational complexity. At the meantime, the existing machine learning classification recognition methods such as support vector machine and relevance vector machine are capable of various PQD events classification only when multi-class classifiers are combined with two-class classifiers. A new approach combining compressive sensing(CS) and sparse representation is proposed to reduce the dimension of power quality signals and classify power quality disturbances. At first, a unknown-type test sample of power quality signals will be represented as a sparse linear combination of its training samples assembled in an over complete dictionary. The dimensionality reduction and characteristics of the test sample will be achieved with the operation of a lower-dimensional projection using a CS-based measure matrix. The sensing matrix of the test sample will be achieved with the identical CS-based measure matrix operated on the over complete dictionary consisting of training sample sets. Then the test sample characteristics will be expressed as a sparse linear combination of the sensing matrix with the noise effect correction function. The measure matrix and the over complete dictionary involved in sparse representation classification model will be optimized based on compressive sensing theory. A fast algorithm will be used to solve the sparse representation and reveal the PQD type from the test sample. The proposed method of feature extraction and dimension reduction for PQD signals could be extremely efficient and independent of PQD signal characteristics with simple and fast operations. The proposed sparse representation classification method could be commonly used for multi-class PQD applications over combining two-class classifiers among support vector machine classification methods.It could provide a potential way for power quality monitoring and diagnosis in the developing smart grid.

电能质量扰动类别众多不规则性大，兼顾各类扰动特性的时频特征提取方法复杂度高、内存空间消耗大；同时现有的机器学习分类识别方法模型复杂、计算量大，限制了电能质量扰动分类的通用性和快速性。压缩感知为电能质量扰动分类识别问题提供了一个崭新的思路。本项目研究基于压缩感知理论的降维映射测试矩阵，将高维扰动信号信息完备地降维映射为低维特征样本，实现普适简单、易于操作的扰动特征提取过程；通过压缩感知建立测试样本与多类别扰动训练集样本之间的通用型稀疏表示关系，利用稀疏表示向量解中蕴含的类别信息，建立直接面向多分类扰动识别任务的稀疏表示扰动多分类模型；协调、约束和优化稀疏表示模型中的测量矩阵、稀疏变换基、快速稀疏求解等各个环节，引入误差因子修正稀疏表示分类模型，增强系统的自适应性及鲁棒性。本研究可为电能质量的检测与识别提供理论和应用基础，为智能电网模式下海量电能质量数据的监测与在线诊断识别提供新的思路和方法。

本课题针对电能质量多类别扰动信号特征提取和识别方法复杂的问题，从一个崭新的视角研究具备通用性和快速性的压缩感知电能质量扰动分类方法。建立了基于压缩感知和稀疏表示理论电能质量多类扰动信号的扰动识别分类模型，分析了压缩感知降维特征提取、超完备训练样本字典、稀疏表示模型、快速稀疏求解、目标归属类识别等各个环节，在保证系统鲁棒性和稳定性的前提下，实现电能质量多扰动识别系统的高效、快速、高识别率的目的。提出了一种基于判别字典学习的稀疏表示电能质量扰动识别方法，避免了传统电能质量扰动识别方法先信号特征提取再人工智能方法分类识别过程的复杂性和冗余性，可有效减少识别步骤、降低复杂性，抗噪声鲁棒性好，在信噪比20dB以上的噪声环境中电能质量扰动识别准确率达到95%以上。提出了一种基于正则化自适应匹配追踪的电能质量数据重构方法，有效改善传统方法在电能质量信号的采集和压缩方面所面临的资源浪费以及重构性能较差等问题，对感知矩阵中的原子进行一次挑选并且计算相关系数，将挑选出的原子索引值存入至候选集中对原子的数目进行自适应地调节，并运用正则化的处理过程完成支撑集的二次挑选，用步长逐步逼近信号的稀疏度进而准确重构出电能质量原始信号。本项目研究成果为电能质量扰动分类识别问题中普适简单的特征提取方法与通用快速的扰动多分类识别方法奠定了技术基础，为海量电能质量数据的监测与在线诊断识别提供了有效方法，具有重要的理论和应用价值。