联合半参数与优化BiLSTM的BDS-3钟差超短期预报  

Ultra-short-term prediction method for BDS-3 clock offset by combined semi-parametric and improved BiLSTM models

作　　者：潘雄 赵万卓 张龙杰 蔡茂 金丽宏 艾青松 PAN Xiong;ZHAO Wanzhuo;ZHANG Longjie;CAI Mao;JIN Lihong;AI Qingsong(School of Computer Science and Artificial Intelligence,Wuhan Textile University,Wuhan 430200,China;School of Mathematical and Physical Sciences,Wuhan Textile University,Wuhan 430200,China;Changjiang Space Information Technology Engineering Co.Ltd.,Wuhan 430010,China)

机构地区：[1]武汉纺织大学计算机与人工智能学院,武汉430200 [2]武汉纺织大学数理科学学院,武汉430200 [3]长江空间信息技术工程有限公司(武汉),武汉430010

出　　处：《中国惯性技术学报》2024年第10期985-993,共9页Journal of Chinese Inertial Technology

基　　金：国家自然科学基金面上项目(42174010,41874009);湖北省自然科学基金(2023AFB435)。

摘　　要：针对半参数钟差预报模型受核函数和窗宽参数选择影响较大以及双向长短期记忆神经网络(BiLSTM)易发生收敛速度慢、预报结果不稳定的问题,提出了一种联合半参数与优化BiLSTM的北斗3号(BDS-3)钟差超短期预报算法。引入莱维飞行算法对BiLSTM超参数进行优化,将半参数模型和优化后的BiLSTM模型组合,构建组合模型(Semi-LFA-BiLSTM)并应用到BDS-3钟差超短期预报。实验结果表明:在12 h预报中,所提组合模型对氢原子钟的预报精度达到0.2 ns左右,铷原子钟的预报结果均能保持在亚纳秒级,其平均精度相较于二次多项式模型、谱分析模型和BiLSTM模型分别提高81.00%、65.29%和44.74%,在钟差超短期预报中表现出较高性能。In addressing the challenges posed by the significant influence of kernel function and window width parameter selection on semi-parametric clock difference prediction models,as well as the issues of slow convergence and unstable forecasting results in bidirectional long short-term memory neural network(BiLSTM),an ultra-short-term prediction method for Beidou navigation satellite system 3(BDS-3)clock offset by combined semi-parametric and improved BiLSTM models is proposed.The Levy flight algorithm is introduced for optimizing the hyperparameters of BiLSTM.By combining the semi-parametric model with the optimized BiLSTM model,the enhanced composite model(Semi-LFA-BiLSTM)is applied to the short-term clock offset prediction of BDS-3.Experimental results demonstrate that the prediction accuracy of the proposed combined model reaches about 0.2 ns for all orbital hydrogen atomic clocks in the 12-hour prediction,and the prediction results for rubidium atomic clocks can be maintained at the sub-nanosecond level,with the average accuracy improved by 81.00%,65.29%and 44.74%,respectively,compared with the quadratic polynomial model,the spectral analysis model and the BiLSTM model.It shows high performance in clock offset ultra-short-term forecast.

关 键 词：半参数 钟差预报 神经网络 超参数 

分 类 号：P228[天文地球—大地测量学与测量工程]