神经网络在日长变化预报中的应用研究

发布时间：2018-01-14 06:23

本文关键词：神经网络在日长变化预报中的应用研究　出处：《中南大学》2011年硕士论文　论文类型：学位论文

【摘要】：地球自转参数预报对于天文学和测地学的理论研究和实际应用具有重要意义。在地球自转参数的预报中,日长变化的预报是难点。日长变化包含了复杂的非线性因素,已有的科学研究包括对它的线性预报和非线性预报。本文在对已有研究进行学习分析的情况下,提出使用非线性神经网络(包括遗传算法(GA)优化的BP神经网络、广义回归神经网络)对日长变化进行预报,并与已有研究成果进行对比分析,得到一些有益的成果,为日长变化的研究增加了新的方法。本文研究的主要内容包括： (1)分析了BP神经网络用于日长变化的基本方法及其存在的不足之处,提出使用遗传算法优化BP神经网络的初始权值和阈值,防止网络陷入局部极小,将遗传算法优化的神经网络用于日长变化预报,并将预报结果与BP神经网络预报的结果进行比较。 (2)由于GA优化的神经网络对日长变化的预报需要不断循环计算以得到最优的权值和阈值,对于冗长天文数据的预报无疑增加了预报时间,这给获取实时快速的日长变化数据带来了困难。文中尝试使用更为简洁高效的广义回归神经网络(GRNN)对日长变化进行预报,这种网络模型不需要循环迭代,是一种局部寻优算法,不会陷入局部极小,算法容易实现,并将预报结果与Schuh(2002)及EOP PCC(2010)的预报结果进行比较。实验结果表明,将广义回归神经网络用于日长变化的预报是切实可行的。 (3)传统的日长变化预报在选取样本数据时,多为按跨度i(i为间隔,取1、2、3…n)取值,而神经网络模拟的是相关事物之间的相关性,相关性越大,神经网络越能够获得足够的先验信息,从而得到更优的输出结果。对于日长变化数据,距离越近的数据之间的相关性越大,而输入样本采用按跨度i输入的方式势必忽略一些相近数据之间的相关性,损失重要的先验信息。所以文中提出输入样本按连续输入的方式选取,并与输入方式按跨度i的方式选取进行日长变化预报的结果进行比较。结果表明,样本采用按跨度输入的方式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。
[Abstract]:The earth rotation parameter prediction has important significance for theoretical research and practical application of astronomy and geodesy. In the prediction of earth rotation parameters, length of day forecast is difficult. LOD contains complex nonlinear factors, including the scientific research on its linear prediction and nonlinear prediction in this paper. Analysis on the existing research situation, put forward the use of nonlinear neural network (including genetic algorithm (GA) optimized BP neural network, generalized regression neural network) to predict the length of day, and compared with the existing research results, get some useful results on variation of length of day adds new methods.
The main contents of this paper are as follows:
(1) and the shortcomings of the BP neural network method for the analysis of the change in the length of the day, we use genetic algorithm to optimize BP neural network initial weights and thresholds, to prevent the network from falling into local minimum, the neural network optimized by genetic algorithm for the length of day forecast, and compare the prediction results with BP neural the network prediction results.
(2) the optimization of GA neural networks to forecast the variation of length of day loop calculations are needed to obtain the optimal weights and thresholds for lengthy astronomical data forecast will increase the forecast time, this to obtain real-time data changes in length of day is difficult. This paper try to use a more generalized regression neural network simple and efficient (GRNN) to predict the length of day, this network model does not need iteration, is a local optimization algorithm, not into the local minimum, the algorithm is easy to realize, and the prediction results with Schuh (2002) and EOP PCC (2010) compared the prediction results. The experimental results show that the generalized regression neural network is used to forecast the length of day is feasible.
(3) when selecting sample data, the traditional day length variation forecast is mostly based on the span of I (I interval, 1,2,3... The n value, and the neural network simulation) is the correlation between things, the greater the correlation, neural network can obtain sufficient prior information, in order to obtain the output results better. For the change in the length of the day, the closer the relationship between the data and the greater input samples according to the span of the I input the way will ignore some correlations between similar data, loss of priori information. So in this paper the input samples according to the continuous input mode selection, and the input by way of I span was selected for the prediction of LOD were compared. The results showed that the samples according to the input span high prediction accuracy in ultra in the short-term forecasting, the sample with the continuous input mode in the short and medium term prediction accuracy is high.

【学位授予单位】：中南大学
【学位级别】：硕士
【学位授予年份】：2011
【分类号】：TP183;P183.31

【参考文献】