基于Spark云平台的变压器故障并行诊断与分析

发布时间：2018-11-13 08:31

【摘要】：随着智能电网的快速发展,电力行业已经进入了“大数据时代”。变压器是电网平稳运行的关键设备,变压器故障诊断方法能够保证电力系统平稳运行。在电力系统中,利用变压器在线监测技术可以及时发现其故障类型。但是由于监测的数量点多,且在一段时间内会多次获取监测数据,因此,其数据量规模急剧增长。通过对数据挖掘算法并行化,实现对海量的电力变压器监测数据快速分析。Spark是分布式内存计算框架,具有轻量级快速处理、兼容Hadoop生态系统、学习成本低、活跃的社区支持、支持多种语言编程接口等特点,为实现对海量的变压器监测数据并行分析提供了新的研究思路。本文介绍了常见的变压器故障类型,详细介绍了传统以及智能的故障诊断方法,并分析了不同方法的优缺点,提出了基于Spark云平台的电力变压器并行诊断与分析方案。选择Spark机器学习库中朴素贝叶斯方法作为电力变压器故障分类方法,以DGA监测数据作为输入,完成故障并行分类实验。实验结果表明基于Spark并行分类方法在性能方面要明显优于单机环境下分类方法。此外,在研究模糊聚类算法的基础上,利用分布式矩阵和广播变量机制,在Spark平台上实现了并行模糊聚类算法Spark-FCM,扩充了Spark机器学习算法库,并将该算法应用在变压器故障聚类上,通过实验表明,该方法具有良好的可行性。
[Abstract]:With the rapid development of smart grid, the power industry has entered the "big data era." Transformer is the key equipment of power network running smoothly. Transformer fault diagnosis method can ensure the smooth operation of power system. In power system, transformer on-line monitoring technology can be used to find fault types in time. However, because of the large number of monitoring points and the acquisition of monitoring data many times in a period of time, the data volume increases rapidly. By parallelizing the data mining algorithm, the fast analysis of massive power transformer monitoring data is realized. Spark is a distributed memory computing framework, which has the advantages of lightweight and fast processing, compatible with Hadoop ecosystem, low learning cost and active community support. It supports many language programming interfaces and provides a new research idea for parallel analysis of massive transformer monitoring data. This paper introduces common types of transformer faults, introduces traditional and intelligent fault diagnosis methods in detail, analyzes the advantages and disadvantages of different methods, and proposes a parallel diagnosis and analysis scheme for power transformers based on Spark cloud platform. The naive Bayes method in Spark machine learning database is selected as the fault classification method of power transformer, and the DGA monitoring data is used as input to complete the parallel fault classification experiment. The experimental results show that the parallel classification method based on Spark is superior to the classification method in single machine environment. In addition, based on the research of fuzzy clustering algorithm, using distributed matrix and broadcast variable mechanism, the parallel fuzzy clustering algorithm Spark-FCM, is implemented on Spark platform to extend the Spark machine learning algorithm library. The algorithm is applied to transformer fault clustering, and the experimental results show that the method is feasible.
【学位授予单位】：华北电力大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.09;TM41

【参考文献】