基于先验簇复杂回声状态网络的话务量预测  

Traffic forecasting for prior knowledge based clustered complex echo state networks

作　　者：雷苗[1] 彭宇[1] 郭嘉[1] 彭喜元[1] 

机构地区：[1]哈尔滨工业大学自动化测试与控制研究所,哈尔滨150080

出　　处：《仪器仪表学报》2011年第10期2190-2197,共8页Chinese Journal of Scientific Instrument

基　　金：教育部高等学校博士学科点专项科研基金(20092302110013);教育部高等学校新世纪优秀人才基金(NCET-10-0062);中国移动通信集团黑龙江有限公司科技发展项目([2008]1891)资助

摘　　要：提出一种基于先验簇复杂回声状态网络的移动通信话务量预测模型。目前,具有小世界、无尺度等特性的复杂网络已被应用于回声状态网络储备池的构建,并获得了较标准回声状态网络储备池更优的回声状态特性和非线性表达能力。在此基础上,针对具有多周期特性的话务量序列预测问题,以功率谱分析的结果作为先验知识,在复杂回声状态网络储备池中构建具有差异性的多个子簇,以期形成对不同频率成分具有表达能力的功能簇。采用中国移动真实数据测试表明,该方法由于考虑了不同周期因素对预测建模的影响,获得了较标准回声状态网络、均匀簇复杂回声状态网络等方法更高的预测精度;以对比方法中实际效果较好的复杂回声状态网络为参照,该方法在齐齐哈尔、大庆、双鸭山市某小区上预测误差分别下降25%、21%和11%;能够为移动通信网络拥塞、覆盖和干扰等问题提供决策支持。This paper proposes a prior knowledge based clustered complex echo state network（PCCESN） for mobile communication traffic forecasting.In order to reflect some learning mechanisms of real world organization,various complexities such as small-world features and scale-free node degree distribution are introduced to the dynamic reservoir of echo state networks（ESNs）,and superior approximation capabilities over the ESNs is achieved.Our further observation on the real traffic collected by Heilongjiang Co.Ltd.,china mobile communications corporation（CMCC） shows the property of multi-periodicities.Based on this characteristic,in this paper power spectrum is chosen as the prior knowledge for generating multiple functional clusters in the dynamic reservoir of complex echo state network.Experiment results for real traffic show that the forecasting accuracy of the proposed PCCESN model is superior to that of original ESN models and Scale-free Highly clustered echo state network（SHESN）.Compared with SHESN model,the prediction error of the proposed method is reduced by a factor of 25%,21% and 11% for the cells in Qiqihar,Daqing and Shuangyashan,respectively,with the consideration of multi-periodicity factors.The proposed PCCESN method can provide decision-making support for network planning and optimization of mobile network.

关 键 词：话务量预测 先验知识 回声状态网络 复杂网络 功能簇 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]