基于高斯过程的日长变化预报

由于日长(length-of-day,LOD)变化具有复杂的时变特性,传统线性模型如最小二乘外推模型、时间序列分析模型等的预报效果往往不甚理想,所以将一种新型的机器学习算法—高斯过程(Gaussian processes,GP)方法用于LOD变化预报,并将预报结果同利用反向传播神经网络(back propagation neural networks,BPNN)和广义回归神经网络(general regression neural networks,GRNN)的预报结果以及地球定向参数预报比较竞赛(Earth Orientation Parameters Prediction Comparison Campaign,EOP PCC)的预报结果进行对比.结果表明,GP用于LOD变化预报是高效可行的.     

大气角动量函数在日长变化非线性预报中的重要作用

日长变化具有复杂的时变特性,传统的线性时间序列分析方法往往难以取得良好的预报效果.采用非线性人工神经网络技术对日长变化进行预报,网络模型的拓扑结构由最小均方误差法来确定.考虑到日长变化与大气环流运动间的密切关系,在神经网络预报模型中引入轴向大气角动量序列.结果表明,联合日长和大气角动量序列,比起单独采用日长资料,预报精度得到显著的提高.     

跳步自回归模型在中长期日长变化预报中的应用

利用自回归模型进行日长变化中长期预报时,预报精度逐渐降低.跳步自回归模型在中长期的预报中具有良好的预报精度,且具有较好的预测稳定性.因此,尝试采用跳步自回归模型替代自回归模型进行日长变化预报.最后,利用国际地球自转参数与参考系服务(International Earth Rotation and Reference Systems Service,IERS)提供的EOP08 C04日长变化序列进行实验,并分析比较两种模型的预报结果.实验结果表明,跳步自回归模型用于改善自回归模型中长期预报精度是可行有效的.     

经验模式分解在极移超短期预报中的应用

经验模式分解(Empirical Mode Decomposition,简称EMD)是一种数据驱动的自适应非线性时变信号分解方法,可以把数据分解成具有物理意义的模式函数分量.采用EMD对极移序列进行分解,去除序列中的高频信号,然后基于最小二乘外推(Least Squares Extrapolation,简称LSE)和广义回归神经网络(General Regression Neural Network,简称GRNN)的组合模型对去除高频信号的极移序列进行1～10d的超短期预报.实验结果表明:将该模型应用到极移超短期预报具有可行性,预报精度有明显改善.     

基于极限学习机的极移中长期预报

传统的极移预报多是基于最小二乘外推和自回归等线性模型,但极移包含了复杂的非线性成分,线性模型的预报效果往往不甚理想。将一种新型神经网络极限学习机(Extreme Learning Machine, ELM)用于极移中长期预报。首先利用最小二乘外推模型对极移序列进行拟合,获得趋势项外推值,然后采用极限学习机对最小二乘拟合残差进行预报,最终的极移预报值为趋势项外推值与残差预报值之和。将极限学习机的预报结果同反向传播(Back Propagation, BP)神经网络与地球定向参数预报比较活动(Earth Orientation Parameters Prediction Comparison Campaign, EOP PCC)的预报结果进行对比,结果表明,极限学习机用于极移中长期预报是高效可行的。     

两种样本输入方式下基于GRNN的日长变化预报结果的比较

针对广义回归神经网络用于日长变化预报过程中,样本的输入方式对预报结果的影响进行了研究。采用2种输入方式:即样本按不同跨度输入以及按连续输入,对日长变化进行预报。最终证明不同的样本输入方式对日长变化预报精度的影响较大,样本按跨度输入在超短期预报中预报精度较高,样本采用连续输入的方式在短期和中期预报中预报精度较高。     

顾及最小二乘拟合端点效应的日长变化预报

针对日长(Length Of Day,LOD)变化预报中最小二乘(Least Squares,LS)拟合存在端点效应的问题,采用时间序列分析方法对日长变化序列进行端点延拓,形成一个新序列,然后用新序列建立最小二乘模型,最后再结合最小二乘模型和自回归(Autoregressive,AR)模型对原始日长变化序列进行预报。实验结果表明,在日长变化序列两端增加统计延拓数据,能有效减小最小二乘拟合序列的端点畸变,从而提高日长变化的预报精度,尤其对中长期预报精度提高明显。     

日长变化预报中BP神经网络拓扑结构的选择

日长变化的预报具有重要的科学意义和实际应用价值。非线性的人工神经网络技术中的反向传播模型（BP网络）可用于预报日长变化。BP网络的拓扑结构决定了神经网络解决问题的能力，针对不同的问题需要采用不同的网络结构。该文分析了神经网络的拓扑结构算法，选用最小均方误差法确定网络的拓扑结构，并将此应用于日长变化预报。结果表明，该方法是可靠和有效的。     

BDS/GPS星载原子钟的短期钟差预报模型研究

针对BDS(BeiDou Navigation Satellite System)/GPS(Global Positioning System)星载原子钟特性和卫星钟差预报模型研究中存在的若干问题,在介绍4种单一模型(多项式模型(QR)、灰色模型(GM)、时间序列模型(ARMA)和广义回归神经网络模型(GRNN))的基础上,引入了经典权组合模型(CM)和修正经典权组合模型(Modified CM).利用武汉大学卫星导航定位技术研究中心的卫星精密钟差产品对BDS/GPS星载原子钟的短期钟差预报模型进行研究,并对比了不同卫星钟和不同模型的预报效果.试验结果表明:单一模型对于BDS卫星钟(C04(GEO Rb)、C07(IGSO Rb)、C14(MEO Rb))的钟差预报精度跳跃性很大;而对于GPS卫星钟(G04(Block IIA Rb)、G09(Block IIA Cs)、G16(Block IIR Rb)、G31(Block II-M Rb)、G27(Block IIF Rb)、G24(Block IIF Cs))的预报精度变化比较平稳;同一种预报模型应用在不同类型的卫星钟序列中,预报精度差异较大.然而,修正经典权组合模型在保证预报可靠性的前提下提高了预报精度,在一定程度上优于其他模型.     

遗传算法优化的BP神经网络卫星钟差预报

针对BP (Back Propagation)神经网络模型预测卫星钟差中权值和阈值的最优化问题, 提出了基于遗传算法优化的BP神经网络卫星钟差短期预报模型, 给出了遗传算法优化BP神经网络的基本思想、具体方法和实施步骤. 为验证该优化模型的有效性和可行性, 利用北斗卫星导航系统(BeiDou navigation satellite system, BDS)卫星钟差数据进行钟差预报精度分析, 并将其与灰色模型(GM(1,1))和BP神经网络模型预报的结果比较分析. 结果表明: 该模型在短期钟差预报中具有较好的精度, 优于GM(1,1)模型和BP神经网络模型.     

Application of General Regression Neural Network to the Prediction of LOD Change

Traditional methods for predicting the change in length of day (LOD change) are mainly based on some linear models, such as the least square model and autoregression model, etc. However, the LOD change comprises complicated non-linear factors and the prediction effect of the linear models is always not so ideal. Thus, a kind of non-linear neural network — general regression neural network (GRNN) model is tried to make the prediction of the LOD change and the result is compared with the predicted results obtained by taking advantage of the BP (back propagation) neural network model and other models. The comparison result shows that the application of the GRNN to the prediction of the LOD change is highly effective and feasible.     

The Important Role Played by Atmospheric Angular Momentum in the Non-linear Prediction of the Variations in the Length of Day

The variation in the length of day has complicated time-varying characteristics and the traditional method for linear time series analysis is always difficult to obtain good effect of prediction. If the non-linear artificial neural network technique is adopted to predict the variation in the length of day, the topological structure of the network model is determined by the least square error method. By taking into account the close relation between the variation in the length of day and the general circulation of atmosphere, the axial sequence of atmospheric angular momentum is introduced into the forecasting model of neural network. The results show that the forecast accuracy is significantly improved by taking advantage of the combination of the length of day and the atmospheric angular momentum sequence in comparison with the individual adoption of the data of the length of day.     

Application of artificial neural networks to the prediction of dust storms in Northwest China

Artificial neural networks (ANN) are non-linear mapping structures analogous to the functioning of the human brain. In this study, we take the ANN approach to model and predict the occurrence of dust storms in Northwest China, by using a combination of daily mean meteorological measurements and dust storm occurrence. The performance of the ANN model in simulating dust storm occurrences is compared with a stepwise regression model. The correlation coefficients between the observed and the estimated dust storm occurrences obtained from the neural network procedure are found to be significantly higher than those obtained from the regression model with the same input data. The prediction tests show that the ANN models used in this study have the potential of forecasting dust storm occurrence in Northwest China by using conventional meteorological variables.     

Speeding Up Image Reconstruction Methods in Coded Mask γ Cameras Using Neural Networks: Application to the EM Algorithm

When using γ-ray coded-mask cameras, one does not get a direct image as in classical optical cameras but the correlation of the mask response with the source. Therefore the data must be mathematically treated in order to reconstruct the original sky sources. Generally this reconstruction is based on linear methods, such as correlating the detector plane with a reconstruction array, or non-linear ones such as iterative or maximization methods (i.e. the EM algorithm). The latter have a better performance but they increase the computational complexity by taking a lot of time to reconstruct an image. Here we present a method for speeding up such kind of algorithms by making use of a neural network with a back-propagation learning rule.     

Atmospheric and Oceanic Excitations to LOD Change on Quasi-biennial Tune Scales

We use wavelet transform to study the time series of the Earth's rotation rate (length-of-day, LOD), the axial components of atmospheric angular momentum (AAM) and oceanic angular momentum (OAM) in the period 1962-2005, and discuss the quasi-biennial oscillations (QBO) of LOD change. The results show that the QBO of LOD change varies remarkably in amplitude and phase. It was weak before 1978, then became much stronger and reached maximum values during the strong El Nino events in around 1983 and 1997. Results from analyzing the axial AAM indicate that the QBO signals in axial AAM are extremely consistent with the QBOs of LOD change. During 1963-2003, the QBO variance in the axial AAM can explain about 99.0% of that of the LOD, in other words, all QBO signals of LOD change are almost excited by the axial AAM, while the weak QBO signals of the axial OAM are quite different from those of the LOD and the axial AAM in both time-dependent characteristics and magnitudes. The combined effects of the axial AAM and OAM can explain about 99.1% of the variance of QBO in LOD change during this period.     

50-day Oscillation of Length-of-day Change

Variations of Earth’s rotation rate (length-of-day, LOD) occur over a wide range of time scales from a few hours to the geological age. Studies showed that the 50-day fluctuation exists in LOD change. In the present paper, the authors use wavelet technique to study the 50-day oscillation in LOD series. Temporal variations of the oscillation are presented in this work. After analyzing the axial component of atmospheric angular momentum (AAM) and oceanic angular momentum (OAM), the 50-day periodic signal is also found in atmospheric and oceanic motion with remarkable time-variation. Meanwhile, the 50-day oscillation of the axial AAM is in good consistence with that of LOD change. This suggests that the 50-day oscillation of LOD change is mainly excited by the axial AAM. Possible origin of the oscillation for Earth system is discussed in the end of this paper.     

基于线性数据重构的天线动力学模型辨识方法

射电望远镜天线伺服控制系统中的非线性特性, 对系统动力学特性辨识有着显著的影响, 会提高辨识难度, 增加辨识模型的复杂程度. 系统非线性特性的测量与补偿也会增加系统辨识工作量. 针对上述问题, 提出了一种基于非线性采样数据的线性重构方法, 用于动力学特性建模. 通过提取原采样数据的相位与幅值, 对受到噪声与非线性畸变影响的系统采样数据进行线性重构, 降低待辨识模型的复杂度. 搭建了半实物实验平台, 以平台实际采样为基础, 重构线性数据, 利用奇异值法与自回归神经网络评估并辨识平台动力学模型. 实验结果表明, 建模数据奇异值拐点从100阶下降至40阶, 仅用10个神经网络节点200次训练即实现了模型辨识.     

Speeding Up Image Reconstruction Methods in Coded Mask γ Cameras Using Neural Networks: Application to the EM Algorithm

When using γ-ray coded-mask cameras, one does not get a direct image as in classical optical cameras but the correlation of the mask response with the source. Therefore the data must be mathematically treated in order to reconstruct the original sky sources. Generally this reconstruction is based on linear methods, such as correlating the detector plane with a reconstruction array, or non-linear ones such as iterative or maximization methods (i.e. the EM algorithm). The latter have a better performance but they increase the computational complexity by taking a lot of time to reconstruct an image. Here we present a method for speeding up such kind of algorithms by making use of a neural network with a back-propagation learning rule. This revised version was published online in July 2006 with corrections to the Cover Date.