【作者】 孔英会；

【导师】 苑津莎；

【作者基本信息】 华北电力大学（河北） ， 电工理论与新技术， 2009， 博士

【摘要】 数据流管理技术的研究已成为信息处理和数据库领域的热点和前沿,数据流管理技术可以为实时信息处理和分析提供有效的支持。随着电网规模的扩大和电力自动化程度的提高,电力系统运行产生的大量实时数据形成数据流,这些数据流中隐含着与设备故障及系统稳定等有关的重要信息。研究数据流技术及其在电力信息处理中的应用,选题具有重要的理论价值与实际价值。论文针对数据流处理的关键技术及在电力信息处理中的应用展开研究,主要内容及成果如下:根据数据流处理实时性的需求建立了通用的数据流处理模型,该模型利用滑动窗口实现最近数据的连续处理,利用概要数据结构和数据流处理算法满足不同的处理需求。论文对离散小波变换概要构建和增量更新算法进行了研究,提出了改进的小波分解增量更新算法,既提高了处理的精度,又减少了运行时间。提出了一种新的基于参数估计的概要构建方法,可满足系统辨识的应用需求,为数据流概要构建提供了新的思路;基于时序特征与参数估计的变压器故障诊断的应用实例说明了所提出概要构建方法的有效性。对数据流连续查询与数据流异常检测技术要解决的问题进行了分析,对基于移动小波树的数据流异常检测算法进行了研究。针对移动小波树异常检测算法存在的不足,提出了改进的基于移动小波树的数据流异常检测算法,通过构建二分检测单调搜索空间和实时更新移动小波树,实现了快速的检测。以电压骤降检测为例进行实验仿真,实验结果表明,提出的异常检测算法具有较小的时间开销,满足实时性要求,为电压暂降事件检测提供了一种新的方法。将传统数据挖掘算法与数据流处理思想相结合,实现了数据流的聚类和分类。并根据分时电价的需求,提出了基于数据流挖掘的用电负荷分类方法;根据在线电能质量扰动识别的需求,提出了基于滑动窗口和增量小波分解的在线电能质量扰动识别方法,可以为电能质量的改善提供决策支持。将传统预测方法与数据流处理的思想相结合,提出了基于小波分解和最小二乘支持向量机的数据流预测方法,通过对不同分解尺度的信号分别建立预测模型保证了预测的精度,采用滑动窗口实现了数据流的连续预测,利用增量小波分解和在线最小二乘支持向量机预测算法减少了运行时间,利用电力负荷数据和发电机功角数据进行仿真实验,实验结果表明,基于增量小波分解与最小二乘支持向量机的数据流预测方法在预测精度和处理时间上都有较好的优势,可用于超短期电力负荷预测与电力系统暂态稳定性预测。更多还原

【Abstract】 Data Stream Management techniques have become hot spots in the fields of information processing and database, it can provide effective support for real-time information processing and analysis. With the expansion of power grids and the raising of automation level, large amount of real-time data generated by the power system operation forms data streams, these data streams include important information related with equipment failures and system stability. Study on data stream techniques and its application in electric power information processing has important theoretical value and practical value. The main content of our research includes:A general data stream processing model is established according to the real-time requirement for data stream processing, it can be used to process recent data using sliding window, and can be suitable for different processing needs by way of synopsis data structure and data stream processing algorithm. How to establish a synopsis data structure using wavelet transforms and update the results incrementally is investigated. An improved incremental update algorithm for wavelet decomposition is proposed, which can improve the handling accuracy and reduce the response time.A new method for constructing a synopsis data structure based on parameters estimation is proposed to meet the need of system identification applications, which provide a new way for the construction of data streams synopsis. A simulation experiment of the transformer fault diagnosis using temporal characteristics and parameters estimation are finished, the experimental results show effectiveness of proposed method.The issues of continuous query and anomaly detection in data stream and related problems are analyzed. The anomaly detection algorithm based on shifted wavelet tree in data streams have been studied, and an improved algorithm was proposed. For the improved algorithm of shifted wavelet tree, monotonous search space was constructed for binary detection which has improved the detection efficiency; and the incremental algorithm of updating Wavelet tree is use to reduce response time. A simulation experiment of detecting voltage sag was finished, the experimental results show that the anomaly detection algorithm has low requirements in processing time and has high detection accuracy, it provides a new approach to real-time voltage sag detection.Clustering and classification of data streams were implemented by combining traditional data mining algorithms and the idea of data stream processing. A method of load classification for power distribution transformer based on data stream mining is proposed which can meet the need of TOU( time-of-use) electricity price, and a way of identifying power quality disturbances online using data stream mining is proposed which can provide decision support for the power quality improvement.A data stream prediction method is proposed by combining traditional prediction methods and the idea of data stream processing. In this method, wavelet decomposition and least squares support vector machines (LS-SVM) are combined to ensure the accuracy of the prediction, sliding window model in data stream processing is used to follow the data changing, and incremental algorithms for wavelet decomposition and online LS-SVM are used to save time. Simulation experiment using real power load dataset and the generator power angle dataset in the transient stability analysis of power system proves the effectiveness of the proposed method. This method can be used to ultra-short-term load prediction and power system transient stability prediction.更多还原