传感器网络环境监测时间序列数据的高斯过程建模与多步预测  

Gaussian process modeling and multi-step prediction for time series data in wireless sensor network environmental monitoring

作　　者：陈艳[1,2] 王子健[1] 赵泽[1] 李栋[1] 崔莉[1] 

机构地区：[1]中国科学院计算技术研究所,北京100190 [2]中国科学院大学,北京100190

出　　处：《通信学报》2015年第10期252-262,共11页Journal on Communications

基　　金：中国科学院战略性先导科技专项基金资助项目(XDA06010403);国家国际科技合作专项基金资助项目(2013DFA10690);国家自然科学基金资助项目(61100179;61202412)~~

摘　　要：针对传感网环境监测应用采集的时间序列数据,提出了一种新的基于高斯过程模型的多步预测方法,实现了对未来时刻的环境监测数据的预测。高斯过程模型通过核函数描述数据的特性,通过对环境监测数据的经验模态分解,以及对其内在物理特性的分析,构建了针对环境监测数据的高斯过程核函数,实现了对数据变化模式的描述。在基于3个数据集的5个种类、20 000多个环境监测数据上进行了性能对比实验,结果表明,与对比预测方法相比,提出的高斯过程多步预测方法对未来时刻的环境监测数据的平均预测精度可以提高20%,可以应用于环境参数未来趋势分析、异常环境事件预警等场景。For time series data collected from WSN environmental monitoring applications, a novel multi-step prediction method based on Gaussian process model was proposed. The method could make prediction for future environmental monitoring data. Kernel functions were used to describe data properties in the Gaussian process model. Kernel functions for environmental monitoring data were constructed through the EMD（empirical mode decomposition） technique and analysis of data inherent physical properties. And the constructed kernel functions were capable of describing the data change mode. Extensive experiments for multi-step prediction performance comparison test were performed on three kinds of data sets using over 20 000 environmental monitoring data records. Experimental results show that the average prediction accuracy of the Gaussian process multi-step prediction method can be increased by 20% than compared prediction methods. The prediction method can be applied to future environmental parameters trend analysis, early warning for abnormal environmental events and other scenes.

关 键 词：传感网环境监测 时间序列 高斯过程 多步预测 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]