一种基于小波大纲的数据流在线预测方法

Online prediction method for data streams based on wavelet synopses

描述了一种基于时间序列数据流大纲的预测框架,提出了构建具有有效降噪效果的小波大纲的方法,可根据背景噪声而分层自适应设置去噪(保留)阈值。并且在这种小波大纲的基础上实现了多尺度概要的分析和预测方法,能够分析动态变化的高频数据流的趋势、拐点、周期、方差的变化,用来为时间序列数据流提供实时的注解。在实际电力负荷数据上的仿真实验证明这种方法可以提供快速的精确的近似预测。

Several studies in recent years were demonstrated that wavelets can be efficiently used to compress large quantities of data down to compact wavelet synopses and provided fast and fairly accurate approximate answers to queries. In this paper author presented a wavelet synopses and prediction framework to analyze dynamic high-frequency data streams. A novel construction method for wavelet synopses provided with efficient De-noise ability was proposed. Its varied threshold schema for every decomposition level could adapt itself to the change of background noise. Based on this wavelet synopsis, a multi-scale prediction and analysis method for summarization was used to separate out the trend, turning points, cyclical fluctuations and autocorrelational effects etc. This framework was used to provide annotation for time series data streams at real-time. Experimental results with real power load datasets demonstrate that our approach achieves improved velocity and accuracy to approximate prediction queries when compared to existing techniques.