组合式深度学习的电离层TEC短期预报模型  

Combined Weighted Deep Learning for Ionospheric TEC Short-term Prediction

作　　者：韦律权 黎峻宇 刘立龙 黄良珂 杨芸珍 魏朋志 WEl Lyuquan;LI Junyu;LlU Lilong;HUANG Liangke;YANG Yunzhen;WEl Pengzhi(Guangxi Vocational College of Water Resources and Electric Pover,Nanning 530023,China;School of Geomaties and Geoinformation,Guilin University ofTechnology,Guilin 541004,China;Guangxi Key Laboratory of Spatial Information and Mapping,Guilin 541004,China)

机构地区：[1]广西水利电力职业技术学院,广西南宁530023 [2]桂林理工大学测绘地理信息学院,广西桂林541004 [3]广西空间信息与测绘重点实验室,广西桂林541004

出　　处：《测绘科学技术学报》2024年第4期369-374,共6页Journal of Geomatics Science and Technology

基　　金：国家自然科学基金项目(42064002);广西科技计划项目(2020GXNSFBA297145);2024年度广西高校中青年教师科研基础能力提升项目(2024KYZD03)。

摘　　要：针对磁暴期间电离层总电子含量TEC异常扰动导致预报精度大幅降低的问题,提出基于强化学习的Q学习算法,对遗传算法优化BP神经网络模型和长短时记忆网络模型进行优化组合,进而建立了一种组合式深度学习的电离层TEC预报模型。分别利用组合模型、两个单一模型对CODE提供的中国地区TEC数据进行3 d预报。结果表明,在不同磁暴等级(强、中、弱、无)下,组合模型预报的平均相对精度分别为95.9%、95.7%、92.6%和95.3%,较两个单一模型平均提高了约6%;其中预报残差小于1 TECu的占比分别达到60%、59%、76%和98%,相比两个单一模型平均提升了约27%。Aiming at the problem that the prediction accuracy is greatly reduced due to the abnormal disturbance of ionospheric Total Electron Content(TEC) during magnetic storm, a Q-Learning algorithm based on reinforcement learning is proposed to optimize the combination of Genetic Algorithm optimized BP neural network model and long-term and short-term memory network model, and then a combined deep learning ionospheric TEC prediction model is established. The combined model and two single models are used to forecast the TEC data in China provided by CODE for three days. The results show that under different levels of magnetic storms(strong, medium, weak, and none), the average relative accuracies of the combined model forecast for three days are 95.9%, 95.7%, 92.6%, and 95.3%, respectively, which is about 6% higher than these of the two single models. Among them, the proportion of forecast residuals less than 1 TECu reaches 60%, 59%, 76% and 98%, which is an average increase of about 27% compared with these of the two single models.

关 键 词：电离层 Q学习 遗传算法改进BP神经网络 长短时记忆网络 组合模型 预报模型 

分 类 号：P228[天文地球—大地测量学与测量工程]