摘要
近年来,基于深度学习技术的长短期记忆(long short-term memory,LSTM)网络相关预报算法在空间天气的预测方面得到广泛应用,但存在预测误差随时间堆叠的缺陷,因此只能进行有限的短期预测。为解决这一问题,文章将太阳风参数、太阳黑子数、地磁活动水平指数Ap以及磁暴环电流指数Dst作为预报因子加入模型,建立一个基于LSTM和深度神经网络(deep neural networks,DNN)的混合模型来进行全球电离层TEC map的中短期预报。该模型可以明显减小时间递增对预测误差的影响。测试结果表明,相较于单独的LSTM模型,LSTM-DNN混合模型对24 h电离层预报准确率相近,对48 h电离层预报平均相对精度(RA)由79.30%提升到81.18%,对144 h电离层预报平均相对精度由64.97%提升到77.64%。
The Long Short-Term Memory(LSTM)related prediction algorithm based on the deep learning technology is widely used in the space weather prediction in recent years.Due to its shortcoming of stacked prediction errors against time,the LSTM-related prediction algorithms can only be used for limited short-term predictions.To improve its performance in the model,we add the solar wind parameters,the sunspot number,the Ap index representing the geomagnetic activity level,and the Dst index corresponding to the magnetic storm ring current as the predictors,to establish a hybrid model based on the LSTM and the DNN(deep neural networks)for medium and short-term forecasting of the global ionosphere TEC map,which can significantly inhibit the increasing trend of the forecast error over time.Test results show that compared with the LSTM model alone,the LSTM-DNN hybrid model enjoys a similar accuracy for the 24-hour prediction of the ionosphere.The average relative accuracy of the 48-hour ionosphere forecast is increased from 79.30%to 81.18%,and that for the 144-hour ionospheric prediction is increased from 64.97%to 77.64%.
作者
廖文梯
陈洲
赵瑜馨
王劲松
唐荣欣
LIAO Wenti;CHEN Zhou;ZHAO Yuxin;WANG Jinsong;Tang Rongxin(Institute of Space Science and Technology,Nanchang University,Nanchang 330031,China;Beijing Institute of Spacecraft Environment Engineering,Beijing 100094,China;Key Laboratory of Space Weather,National Center of Space Weather,China Meteorological Administration,Beijing 100081,China)
出处
《航天器环境工程》
北大核心
2021年第3期281-286,共6页
Spacecraft Environment Engineering
基金
国家自然科学基金项目(编号:41974183)
国家自然科学基金项目(编号:41774195)。
关键词
电离层预报
总电子含量
长短期记忆
深度神经网络
ionosphere forecast
total electron content(TEC)
long short-term memory(LSTM)
deep neural networks(DNN)