利用深度学习实现Dst指数短期业务预报  

Using Deep Learning to Achieve Short Term Business Forecast of Dst Index

作　　者：牛犇 黄智 NIU Ben;HUANG Zhi(School of Physics and Electronic Engineering,Jiangsu Normal University,Xuzhou 221116)

机构地区：[1]江苏师范大学物理与电子工程学院,徐州221116

出　　处：《空间科学学报》2025年第1期91-101,共11页Chinese Journal of Space Science

基　　金：国家自然科学基金项目(41104096);徐州科技计划项目(KC21159)共同资助。

摘　　要：由太阳活动引发的磁暴事件会导致地球磁场产生剧烈变化,进而影响通信、导航、电力等工程应用系统的服务性能.在空间物理领域通常利用Dst指数表征磁暴强度的变化,本文提出一种基于卷积神经网络(CNN)、门控循环单元(GRU)和长短时记忆网络(LSTM)的磁暴预测模型(C-G-LSTM),能够提前1~6 h预测Dst指数.进一步利用美国航空航天局(National Aeronautics and Space Administration,NASA)提供的2010-2019年Dst指数评估混合深度学习预测模型的性能.结果显示最大均方根误差不超过7.29 nT;最大平均绝对误差不超过5.03 nT,磁暴期间误差有所增大.与已有研究结果相比,本文所提出的模型具有较高精度,且无须提供太阳风温度、太阳风动压以及行星际磁场分量等输入参数,适用于业务预报.Magnetic storm events triggered by solar activity can cause dramatic changes in the Earth’s magnetic field,significantly impacting the performance of systems such as communications,navigation,and power supply.These disturbances can interfere with radio signal propagation,reduce navigation accuracy,and disrupt power transmission networks.Therefore,accurately predicting magnetic storms is crucial for mitigating their effects.In space physics,the Dst index is commonly used to characterize the intensity of magnetic storms.It serves as a vital global indicator of geomagnetic activity.To enhance the prediction of magnetic storms and reduce their adverse effects,an efficient and accurate predictive model is essential.This paper proposes a magnetic storm prediction model based on Convolutional Neural Networks(CNN),Gated Recurrent Units(GRU),and Long Short-Term Memory networks(LSTM),referred to as the C-G-LSTM model.This hybrid model leverages the strengths of CNN,GRU,and LSTM to predict the Dst index 1 to 6 h in advance,providing valuable lead time for responding to potential magnetic storm events.CNNs effectively extract spatial features from input data,while GRUs and LSTMs excel at handling time series data and capturing temporal dependencies.The performance of the C-G-LSTM model was evaluated using Dst index data provided by NASA,covering the period from 2010 to 2019.The results demonstrate that this model performs exceptionally well in predicting the Dst index.Specifically,the maximum Root Mean Square Error(RMSE)does not exceed 7.29 nT,and the Maximum Mean Absolute Error(MAE)does not exceed 5.03 nT.Although errors increase during intense magnetic activity,the model maintains high accuracy.A significant advantage of the C-G-LSTM model is that it does not require additional input parameters such as solar wind temperature,solar wind dynamic pressure,and interplanetary magnetic field components,which are often needed in other models.This makes the C-G-LSTM model more straightforward and practical for operational foreca

关 键 词：DST指数 卷积神经网络 门控循环单元 长短时记忆网络 

分 类 号：P352[天文地球—空间物理学]