摘要
针对太阳活动影响下机器学习模型对电离层总电子含量(TEC)短期预报精度不高的问题,本文提出了一种基于变分模态分解(VMD)、麻雀搜索算法(SSA)、长短期记忆神经网络(LSTM)的组合模型(VMD-SSA-LSTM),以期提高TEC短期预报精度.利用VMD算法对不同时期太阳活动程度影响下的东、西半球TEC格网点数据分解,利用SSA优化LSTM模型,将分解的TEC样本分量及模型最优初始权值和阈值输入到LSTM模型中,将分量预测序列合并重构,得到电离层TEC预测值.实验表明:VMD-SSA-LSTM组合模型在东、西半球太阳活动强烈、适中、较弱时期的均方根误差分别为0.77、0.56、0.69;0.92、0.76、0.73个TECu,平均绝对误差平均值分别为0.69、0.47、0.56;0.79、0.65、0.58个TECu,平均相对精度分别达到94％、94％、93％;93％、91％、91％以上,残差绝对值分布在0～1个TECu的比例均值分别为75.56％、96.11％、85％;74.44％、80.55％、78.33％,较VMD-LSTM、LSTM两种模型预报精度有显著提升.
Abstract
Aiming at the problem of poor short-term forecast accuracy of ionospheric total electron content (TEC) by machine learning models under the influence of solar activity, a combined model based on variational modal decomposition (VMD), sparrow search algorithm (SSA), and long-short-term memory neural network (LSTM) (VMD-SSA-LSTM combined model) is proposed to improve the short-term forecast accuracy of TEC.The VMD algorithm is used to decompose the east-west hemispheric TEC grid point data under the influence of different time solar activity levels, optimize the LSTM model by SSA, input the decomposed TEC sample components and the optimal initial weights and thresholds of the model into the LSTM model, and then merge the component prediction sequences into a reconstructed one to obtain the ionospheric TEC prediction values.The experiments show that the root-mean-square errors of the combined VMD-SSA-LSTM model in the eastern and western hemispheres during the periods of strong, moderate, and weak solar activities are 0.77, 0.56, and 0.69; 0.92, 0.76, and 0.73 TECu, respectively, and the mean absolute errors average 0.69, 0.47, and 0.56; 0.79, 0.65, and 0.58 TECu, the average relative accuracies reached 94％, 94％, 93％; 93％, 91％, 91％ or more, respectively, and the mean value of the proportion of the absolute value of the residuals distributed in 0-1 TECu was 75.56％, 96.11％, 85％; 74.44％, 80.55％, 78.33％, respectively, which is a significant improvement over the forecast accuracy of the two models of VMD-LSTM, LSTM improvement.
维普
万方数据