基于改进的深度置信网络的电离层F2层临界频率预测
|
摘要
提出一种基于深度置信网络(deep belief network,DBN)对本区域未来24 h的电离层临界频率f0F2预测的方法。对选取的数据集进行筛选,生成用于训练和测试的数据集;改进DBN基本单元的结构,以适应对连续型数据特征的提取与学习,再通过实验确定DBN的基本结构;最后利用训练数据集对改进后的网络进行训练,实现对f0F2值的预测。与实测值相比较,改进的DBN具有极佳的预测准确性;与浅层结构BP网络和SVM网络相比,改进的DBN不单克服了浅层结构所固有的问题,更表现出对于连续型数据预测的优异性能,尤其是当预测对象受到高维复杂因素影响时改进的DBN模型依旧能表现出很好的预测性能。
This paper proposed a method which was predicting the ionospheric critical frequency f0 F2 of the future 24 h based on deep belief network( DBN). First,it filtered the data and processed into data sets for training and testing. Secondly,it improved the structure of the basic unit of DBN to adapt to the extraction and learning of continuous data feature,and then determined the basic structure of DBN through experiments. Finally,this paper used the training data set to train the improved network to realize the prediction of f0 F2 value. Compared with the measured values,the improved DBN has excellent prediction accuracy. Compared with the shallow structure of BP network and SVM network,the improved DBN not only overcomes the inherent problems of the shallow structure,but also shows the excellent performance of continuous data prediction,especially when the prediction value is affected by high dimensional complex factors,the improved DBN model can still show good prediction performance.
This paper proposed a method which was predicting the ionospheric critical frequency f0 F2 of the future 24 h based on deep belief network( DBN). First,it filtered the data and processed into data sets for training and testing. Secondly,it improved the structure of the basic unit of DBN to adapt to the extraction and learning of continuous data feature,and then determined the basic structure of DBN through experiments. Finally,this paper used the training data set to train the improved network to realize the prediction of f0 F2 value. Compared with the measured values,the improved DBN has excellent prediction accuracy. Compared with the shallow structure of BP network and SVM network,the improved DBN not only overcomes the inherent problems of the shallow structure,but also shows the excellent performance of continuous data prediction,especially when the prediction value is affected by high dimensional complex factors,the improved DBN model can still show good prediction performance.
引文
[1]徐彤,吴健,吴振森,等.基于电离层暴时f0F2经验模型Kalman滤波短期预报[J].空间科学学报,2009,29(2):202-207.
[2]曹红艳,孙宪儒.新版亚大地区F2电离层频率预测方法[J].空间科学学报,2009,29(5):502-507.
[3]鲁转侠,曹红艳,冯静.“新版亚大地区F2电离层预测”方法数据验证[J].中国电子科学研究院学报,2011,6(1):59-63.
[4]刘瑞源,刘顺林,徐中华,等.自相关分析法在中国电离层短期预报中的应用[J].科学通报,2005,50(24):2781-2785.
[5]奚迪龙,李建儒,刘玉梅,等.一种电离层f0F2和M(3000)F2长期预测新方法[J].电波科学学报,2008,23(5):946-949.
[6]Barkhatova O M,LevitinАЕ.Neural network technique of layer F2critical frequency above station Gakona(HAARP)at the account of near-earth space parameters and geomagnetic disturbance[C]//Proc of the 16th Annual Seminar Physics of Auroral Phenomena.[S.l.]:Kola Science Centre,Russian Academy of Science,2008:137-140.
[7]金会彬,张灿,秦磊.基于神经网络的电离层F2层临界频率预测方法[J].中国科学院研究生院学报,2008,25(3):403-407.
[8]Hinton G E,Osindero S,Teh Y W.A fast learning algorithm for deep belief nets[J].Neural Computation,2006,18(7):1527-1554.
[9]乔俊飞,潘广源,韩红桂.一种连续型深度信念网的设计与应用[J].自动化学报,2015,41(12):2138-2146.
[10]Spiliopoulou A.Investigation of Deep CRBM networks in modeling sequential data[D].[S.l.]:University of Edinburgh,2008.
[11]Chen H,Murray A.Continuous restricted boltzmann machine with an implementable training algorithm[J].Image Signal Processing,2003,150(3):153-158.
[12]Marks T K,Movellan J R.Diffusion networks,products of experts,and factor analysis[R].[S.l.]:UCSD MPLab,2001.
[13]Hopfield J J.Neurons with graded response have collective computational properties like those of two-state neurons[J].Proceeding of the National Academy of Sciences of the United States of America,1984,81(10):3088-3092.
[14]Frey B J.Continuous sigmoidal belief networks trained using slice sampling[C]//Advance in Neural Information Processing Systems.Cambridge:MIT Press,1997:452-458.
[15]潘广源,柴伟,乔俊飞.DBN网络的深度确定方法[J].控制与决策,2015,30(2):256-260.
[16]陈春,吴振森,孙树计,等.利用神经网络预报中国地区电离层f0F2[J].空间科学学报,2011,31(3):304-310.
[2]曹红艳,孙宪儒.新版亚大地区F2电离层频率预测方法[J].空间科学学报,2009,29(5):502-507.
[3]鲁转侠,曹红艳,冯静.“新版亚大地区F2电离层预测”方法数据验证[J].中国电子科学研究院学报,2011,6(1):59-63.
[4]刘瑞源,刘顺林,徐中华,等.自相关分析法在中国电离层短期预报中的应用[J].科学通报,2005,50(24):2781-2785.
[5]奚迪龙,李建儒,刘玉梅,等.一种电离层f0F2和M(3000)F2长期预测新方法[J].电波科学学报,2008,23(5):946-949.
[6]Barkhatova O M,LevitinАЕ.Neural network technique of layer F2critical frequency above station Gakona(HAARP)at the account of near-earth space parameters and geomagnetic disturbance[C]//Proc of the 16th Annual Seminar Physics of Auroral Phenomena.[S.l.]:Kola Science Centre,Russian Academy of Science,2008:137-140.
[7]金会彬,张灿,秦磊.基于神经网络的电离层F2层临界频率预测方法[J].中国科学院研究生院学报,2008,25(3):403-407.
[8]Hinton G E,Osindero S,Teh Y W.A fast learning algorithm for deep belief nets[J].Neural Computation,2006,18(7):1527-1554.
[9]乔俊飞,潘广源,韩红桂.一种连续型深度信念网的设计与应用[J].自动化学报,2015,41(12):2138-2146.
[10]Spiliopoulou A.Investigation of Deep CRBM networks in modeling sequential data[D].[S.l.]:University of Edinburgh,2008.
[11]Chen H,Murray A.Continuous restricted boltzmann machine with an implementable training algorithm[J].Image Signal Processing,2003,150(3):153-158.
[12]Marks T K,Movellan J R.Diffusion networks,products of experts,and factor analysis[R].[S.l.]:UCSD MPLab,2001.
[13]Hopfield J J.Neurons with graded response have collective computational properties like those of two-state neurons[J].Proceeding of the National Academy of Sciences of the United States of America,1984,81(10):3088-3092.
[14]Frey B J.Continuous sigmoidal belief networks trained using slice sampling[C]//Advance in Neural Information Processing Systems.Cambridge:MIT Press,1997:452-458.
[15]潘广源,柴伟,乔俊飞.DBN网络的深度确定方法[J].控制与决策,2015,30(2):256-260.
[16]陈春,吴振森,孙树计,等.利用神经网络预报中国地区电离层f0F2[J].空间科学学报,2011,31(3):304-310.