提出一种基于深度学习的电离层f0F2短期预报方法,通过采用注意力机制的双向长短期记忆网络(Bidirectional Long Short-Term Memory Model With Attention Mechanism,BiLSTM-Attention)算法,结合前7天垂测站电离层f0F2观测...提出一种基于深度学习的电离层f0F2短期预报方法,通过采用注意力机制的双向长短期记忆网络(Bidirectional Long Short-Term Memory Model With Attention Mechanism,BiLSTM-Attention)算法,结合前7天垂测站电离层f0F2观测值、世界时、太阳活动指数及地磁活动指数作为输入,实现了中国区域电离层f0F2的预报.模型对比分析结果表明:低纬度台站的预报误差显著高于中纬度台站,BiLSTM-Attention模型表现最优,长短期记忆网络(LSTM)模型次之,与国际参考电离层模型(IRI)相比,BiLSTM-Attention模型的均方根误差(RMSE)降低了44.2%,平均绝对误差(MAE)降低47%,而决定系数(R2)提升21.3%;磁暴期间,BiLSTM-Attention模型成功捕捉中国区域电离层负暴效应(f0F2下降),与观测值非常一致,而IRI模型开启暴时模式后,f0F2预测值与实际观测值之间依然存在一定偏差;随着预报时间从1 h增加至24 h,模型预报误差呈系统性上升趋势,RMSE从0.99 MHz增至2.05 MHz,MAE从0.69 MHz升至1.57 MHz,R2则由0.93减至0.75.相关研究为空间天气预警及短波通信系统优化提供了高精度电离层参数的预报支撑.展开更多
The Rate of TEC Index(ROTI),a crucial indicator of fluctuations in the ionospheric Total Electron Content(TEC),has long served as an effective proxy for characterizing ionospheric activity.However,ROTI sequences exhib...The Rate of TEC Index(ROTI),a crucial indicator of fluctuations in the ionospheric Total Electron Content(TEC),has long served as an effective proxy for characterizing ionospheric activity.However,ROTI sequences exhibit significant nonstationary and nonlinear characteristics,which intensify during substorms,making high-accuracy prediction challenging for traditional models.To address these complications,we propose a hybrid forecasting model,named BWO-VMD-LSTM,that integrates the Beluga Whale Optimization(BWO)algorithm,Variational Mode Decomposition(VMD),and Long Short-Term Memory(LSTM)networks.Our methodology first employs BWO to adaptively optimize the key parameters of VMD,achieving an optimal decomposition configuration.The optimized VMD then decomposes the original non-stationary ROTI sequence into several relatively stable Intrinsic Mode Functions(IMFs).Subsequently,an LSTM model is constructed to independently forecast each IMF.Finally,the predictions of all components are reconstructed to produce the final ROTI forecast.Experimental results demonstrate that our model outperforms baseline models under various geomagnetic conditions(including quiet,normal,and substorm periods),showing significant improvements in key metrics such as Mean Absolute Error(MAE),Root Mean Square Error(RMSE),and Mean Absolute Percentage Error(MAPE),thereby exhibiting superior prediction accuracy and stability.This study presents an effective and robust tool for high-precision ROTI forecasting,with promising potential for enhancing space weather monitoring and scintillation warning services.展开更多
The Triple Ionosphere Photometer(TRIPM)is a scientific payload aboard the Fengyun-3E(FY-3E)satellite,which operates in a dawn−dusk orbit.It is primarily designed for nadir observations of airglow emissions at OI 135.6...The Triple Ionosphere Photometer(TRIPM)is a scientific payload aboard the Fengyun-3E(FY-3E)satellite,which operates in a dawn−dusk orbit.It is primarily designed for nadir observations of airglow emissions at OI 135.6 nm and N2Lyman-Birge-Hopfield(LBH)bands.Due to the satellite’s dawn−dusk orbital characteristics,most of TRIPM’s field of view remains in a semi-illuminated condition.Therefore,compared with airglow data of the same bands acquired under purely daytime or nighttime conditions,applying TRIPM data poses greater challenges.This study presents the first attempt to use TRIPM data for retrieving solar extreme ultraviolet(EUV)flux.Our results demonstrate that by utilizing TRIPM data in regions where photoelectron excitation dominates as the primary radiation source,the solar EUV flux(denoted as QEUV)can be retrieved.Comparisons with data from the SOHO/SEM instrument reveal excellent consistency,with a seasonal correlation coefficient(R)of at least 0.95.This work thus offers a new avenue for solar EUV flux acquisition and expands the application range of TRIPM data.展开更多
现有的电离层总电子含量(Total Electron Content,TEC)时空预测模型主要以堆叠ConvLSTM单元及其变体为主.这种依赖于ConvLSTM的TEC时空预测模型在捕捉局部时空依赖性的时候比较有效.但由于缺乏存储长距离空间记忆的单元,致使长距离的TE...现有的电离层总电子含量(Total Electron Content,TEC)时空预测模型主要以堆叠ConvLSTM单元及其变体为主.这种依赖于ConvLSTM的TEC时空预测模型在捕捉局部时空依赖性的时候比较有效.但由于缺乏存储长距离空间记忆的单元,致使长距离的TEC空间特征依赖难以被ConvLSTM及其变体捕捉.为解决该问题,本文提出了一个基于自注意力记忆卷积长短期记忆网络的电离层TEC时空预测模型SA-ConvLSTM,该模型在具有短期记忆依赖的ConvLSTM基础上,增加了具有长距离记忆依赖的自注意力记忆(self-attention memory,SAM)模块,以便在TEC时空预测中同时兼顾短期记忆和长距离记忆.为了验证SA-ConvLSTM的性能,本文在12.5°S—87.5°N,25°E—180°E区域内选择3年太阳活动高年和3年太阳活动低年的TEC网格数据,在该数据上,将SA-ConvLSTM与目前主流的TEC时空预测模型ConvGRU、ConvLSTM、PredRNN、Residual Attention-BiConvLSTM及CODE提供的电离层预测产品C1PG进行了对比.结果表明,与C1PG、ConvGRU、ConvLSTM、PredRNN和Residual Attention-BiConvLSTM相比,SA-ConvLSTM的RMSE在太阳活动高年分别降低了6.58%、3.89%、5.79%、1.44%、1.21%;在太阳活动低年分别降低了13.42%、10.26%、11.40%、3.20%、4.37%.此外,本文还在不同月份和纬度区域情况下进行了对比,结果表明,在绝大多数月份和绝大多数纬度区域内,SA-ConvLSTM的预测性能更好.最后本文选取了两次磁暴事件来验证SA-ConvLSTM在极端情况下的预测能力.结果表明,SA-ConvLSTM在磁暴的大多数阶段均优于对比模型.展开更多
The Vertical Total Electron Content(VTEC)of the ionosphere is a crucial parameter for describing the distribution and dynamic changes within the ionosphere.The study utilizes Dual Hybrid Attentional UNet(DHA-UNet)mode...The Vertical Total Electron Content(VTEC)of the ionosphere is a crucial parameter for describing the distribution and dynamic changes within the ionosphere.The study utilizes Dual Hybrid Attentional UNet(DHA-UNet)model to achieve higher forecasting performance for global VTEC predictions under the condition of data acquisition delays.Initially,this study uses the first Hybrid Attentional UNet(HA-UNet)model to predict the intermediate missing data.The missing data are caused by delays in data processing,making the Global Ionosphere Map(GIM)for the current day unavailable.Subsequently,the predicted results from the first HA-UNet model are concatenated with the input data to serve as the input data for the second HA-UNet model,yielding the final prediction results.The performance of DHA-UNet model is then evaluated under varying solar and geomagnetic activity conditions.Evaluation results demonstrate that the DHA-UNet model exhibits higher forecasting accuracy and stability compared to commonly used temporal and spatiotemporal forecasting models.Compared to CODG VTEC,the DHA-UNet model achieves Mean Absolute Error(MAE)values of 2.60 TECU,3.07 TECU,3.78 TECU,and 6.45TECU during quiet,weak,moderate,and strong geomagnetic storm periods,respectively,in years of high solar activity.In years of low solar activity,the model achieves MAE values of 1.00 TECU,1.15 TECU,and 1.54 TECU during quiet,weak,and moderate geomagnetic storm periods,respectively.Even during strong geomagnetic storms,55%of the residuals from the DHA-UNet model fall within the-5.0 TECU to 5.0 TECU range,surpassing other commonly used models.Compared to the C1PG forecasting product,the DHA-UNet model shows particularly notable improvements in accuracy during the spring and winter seasons,as well as in mid-to high-latitude regions.展开更多
摘要提出一种基于深度学习的电离层f0F2短期预报方法,通过采用注意力机制的双向长短期记忆网络(Bidirectional Long Short-Term Memory Model With Attention Mechanism,BiLSTM-Attention)算法,结合前7天垂测站电离层f0F2观测值、世界时、太阳活动指数及地磁活动指数作为输入,实现了中国区域电离层f0F2的预报.模型对比分析结果表明:低纬度台站的预报误差显著高于中纬度台站,BiLSTM-Attention模型表现最优,长短期记忆网络(LSTM)模型次之,与国际参考电离层模型(IRI)相比,BiLSTM-Attention模型的均方根误差(RMSE)降低了44.2%,平均绝对误差(MAE)降低47%,而决定系数(R2)提升21.3%;磁暴期间,BiLSTM-Attention模型成功捕捉中国区域电离层负暴效应(f0F2下降),与观测值非常一致,而IRI模型开启暴时模式后,f0F2预测值与实际观测值之间依然存在一定偏差;随着预报时间从1 h增加至24 h,模型预报误差呈系统性上升趋势,RMSE从0.99 MHz增至2.05 MHz,MAE从0.69 MHz升至1.57 MHz,R2则由0.93减至0.75.相关研究为空间天气预警及短波通信系统优化提供了高精度电离层参数的预报支撑.
基金supported by the National Key R&D Program of China(2022YFC2807205)the National Natural Science Foundation of China(Grants 42374208,42120104003)+2 种基金the Industry-UniversityResearch Cooperation Fund of the 8th Research Institute of China Aerospace Science and Technology Corporation(SAST2023-025)the Chinese Meridian Projectthe High Performance Computing Center of Nanjing University of Information Science&Technology for their support of this work。
摘要The Rate of TEC Index(ROTI),a crucial indicator of fluctuations in the ionospheric Total Electron Content(TEC),has long served as an effective proxy for characterizing ionospheric activity.However,ROTI sequences exhibit significant nonstationary and nonlinear characteristics,which intensify during substorms,making high-accuracy prediction challenging for traditional models.To address these complications,we propose a hybrid forecasting model,named BWO-VMD-LSTM,that integrates the Beluga Whale Optimization(BWO)algorithm,Variational Mode Decomposition(VMD),and Long Short-Term Memory(LSTM)networks.Our methodology first employs BWO to adaptively optimize the key parameters of VMD,achieving an optimal decomposition configuration.The optimized VMD then decomposes the original non-stationary ROTI sequence into several relatively stable Intrinsic Mode Functions(IMFs).Subsequently,an LSTM model is constructed to independently forecast each IMF.Finally,the predictions of all components are reconstructed to produce the final ROTI forecast.Experimental results demonstrate that our model outperforms baseline models under various geomagnetic conditions(including quiet,normal,and substorm periods),showing significant improvements in key metrics such as Mean Absolute Error(MAE),Root Mean Square Error(RMSE),and Mean Absolute Percentage Error(MAPE),thereby exhibiting superior prediction accuracy and stability.This study presents an effective and robust tool for high-precision ROTI forecasting,with promising potential for enhancing space weather monitoring and scintillation warning services.
基金supported financially by National Natural Science Foundation of China(Grant No.42174226,42474239)National Key Research and Development Program(2022YFF0503901)China Meteorological Administration‘Ionospheric Forecast and Alerting’Youth Innovation Team(CMA2024QN09).
摘要The Triple Ionosphere Photometer(TRIPM)is a scientific payload aboard the Fengyun-3E(FY-3E)satellite,which operates in a dawn−dusk orbit.It is primarily designed for nadir observations of airglow emissions at OI 135.6 nm and N2Lyman-Birge-Hopfield(LBH)bands.Due to the satellite’s dawn−dusk orbital characteristics,most of TRIPM’s field of view remains in a semi-illuminated condition.Therefore,compared with airglow data of the same bands acquired under purely daytime or nighttime conditions,applying TRIPM data poses greater challenges.This study presents the first attempt to use TRIPM data for retrieving solar extreme ultraviolet(EUV)flux.Our results demonstrate that by utilizing TRIPM data in regions where photoelectron excitation dominates as the primary radiation source,the solar EUV flux(denoted as QEUV)can be retrieved.Comparisons with data from the SOHO/SEM instrument reveal excellent consistency,with a seasonal correlation coefficient(R)of at least 0.95.This work thus offers a new avenue for solar EUV flux acquisition and expands the application range of TRIPM data.
基金funded by the National Key R&D Program of China(No.2022YFB3904402)the National Natural Science Foundation of China(Nos.42474037 and U2233217)。
摘要The Vertical Total Electron Content(VTEC)of the ionosphere is a crucial parameter for describing the distribution and dynamic changes within the ionosphere.The study utilizes Dual Hybrid Attentional UNet(DHA-UNet)model to achieve higher forecasting performance for global VTEC predictions under the condition of data acquisition delays.Initially,this study uses the first Hybrid Attentional UNet(HA-UNet)model to predict the intermediate missing data.The missing data are caused by delays in data processing,making the Global Ionosphere Map(GIM)for the current day unavailable.Subsequently,the predicted results from the first HA-UNet model are concatenated with the input data to serve as the input data for the second HA-UNet model,yielding the final prediction results.The performance of DHA-UNet model is then evaluated under varying solar and geomagnetic activity conditions.Evaluation results demonstrate that the DHA-UNet model exhibits higher forecasting accuracy and stability compared to commonly used temporal and spatiotemporal forecasting models.Compared to CODG VTEC,the DHA-UNet model achieves Mean Absolute Error(MAE)values of 2.60 TECU,3.07 TECU,3.78 TECU,and 6.45TECU during quiet,weak,moderate,and strong geomagnetic storm periods,respectively,in years of high solar activity.In years of low solar activity,the model achieves MAE values of 1.00 TECU,1.15 TECU,and 1.54 TECU during quiet,weak,and moderate geomagnetic storm periods,respectively.Even during strong geomagnetic storms,55%of the residuals from the DHA-UNet model fall within the-5.0 TECU to 5.0 TECU range,surpassing other commonly used models.Compared to the C1PG forecasting product,the DHA-UNet model shows particularly notable improvements in accuracy during the spring and winter seasons,as well as in mid-to high-latitude regions.