基于BiLSTM-Attention的F10.7指数预测模型与中国自主数据集的应用
doi: 10.11728/cjss2024.02.2023-0040 cstr: 32142.14.cjss2024.02.2023-0040
-
闫帅楠：男, 2001年10月出生于江苏省徐州市, 2023年7月本科毕业于江苏科技大学. 现为中国科学院大学硕士研究生在读, 主要研究方向为空间智能信息处理技术, 人工智能与空间灾害性天气事件预报. E-mail: shuainanyan2021@163.com
作者简介:
通讯作者:
Application of F10.7 Index Prediction Model Based on BiLSTM-attention and Chinese Autonomous Dataset
-
摘要: F10.7指数是太阳活动的重要指标, 准确预测F10.7指数有助于预防和缓解太阳活动对无线电通信、导航和卫星通信等领域的影响. 基于F10.7射电流量的特性, 在双向长短时记忆网络(Bidirectional Long Short-Term Memory Network, BiLSTM)基础上融入注意力机制(Attention), 提出了一种基于BiLSTM-Attention的F10.7预报模型. 在加拿大DRAO数据集上其平均绝对误差(MAE)为5.38, 平均绝对百分比误差(MAPE)控制在5%以内, 相关系数(R)高达0.987, 与其他RNN模型相比拥有优越的预测性能. 针对中国廊坊L&S望远镜观测的F10.7数据集, 提出了一种转换平均校准(Conversion Average Calibration, CAC)方法进行数据预处理, 处理后的数据与DRAO数据集具有较高的相关性. 基于该数据集对比分析了RNN系列模型的预报效果, 实验结果表明, BiLSTM-Attention和BiLSTM两种模型在预测F10.7指数方面具有较好的优势, 表现出较好的预测性能和稳定性.Abstract: The F10.7 index is an important indicator of solar activity. Accurate predictions of the F10.7 index can help prevent and mitigate the effects of solar activity on areas such as radio communications, navigation and satellite communications. Based on the properties of the F10.7 radio flux, the prediction model of F10.7 based on BiLSTM-Attention is proposed by incorporating an Attention mechanism on the Bidirectional Long Short-Term Memory Network (BiLSTM). The Mean Absolute Error (MAE) on the Canadian DRAO dataset is 5.38, the Mean Absolute Percentage Error (MAPE) is controlled to within 5% and the correlation coefficient (R) reaches 0.987. It has superior prediction performance compared with other RNN models in both short-term and medium-term prediction. A Conversion Average Calibration (CAC) method is proposed to preprocess the F10.7 data set observed by the Langfang L&S telescope in China. The processed data has high correlation with the DRAO dataset. Based on this dataset the forecasting effectiveness of the RNN series models is compared and analyzed. The experimental results show that both BiLSTM-Attention and BiLSTM models have significant advantages in predicting the F10.7 index and show excellent predictive performance and good stability. The BiLSTM-Attention model has the highest prediction accuracy when forecasting future first-day data, with MAE and MAPE of 11.10 and 8.66, respectively, and the MAPE is always within 15% in the short- and medium-term forecasts. This shows that the proposed model has high generalization ability and can effectively predict the F10.7 data set of DRAO and L&S.
-
图 7 基于BiLSTM-Attention模型的DRAO数据预测值与真实值对比
Figure 7. Comparison of predicted and true values of DRAO data based on BiLSTM-Attention model
图 8 基于BiLSTM-Attention模型的未来27天DRAO数据F10.7指数预测
Figure 8. F10.7 index forecast for the next 27 days based on BiLSTM-Attention model
图 9 基于BiLSTM-Attention模型的L&S数据预测值和真实值对比
Figure 9. Comparison of predicted and true values of L&S data based on BiLSTM-Attention model
图 10 基于BiLSTM-Attention模型的未来27天L&S数据F10.7指数预测曲线
Figure 10. F10.7 Index forecast for the next 27 days based on BiLSTM-Attention model
表 1 参数配置信息
Table 1. Parameter configuration information
参数配置项 值或信息 训练轮次(epoch) 80 学习率 0.001 损失函数 MSE batch_size 8 优化器 RMSprop 
下载: 导出CSV
表 2 各模型在DRAO数据集上多天预测性能比较
Table 2. Comparison of multi-day forecast performance of each model on DRAO dataset
预测天数 模型 Metrics RMSE /sfu MAE/sfu MAPE/ (%) R2 R 1 LSTM 18.50 7.55 7.16 0.808 0.986 BiLSTM 17.46 5.97 5.16 0.828 0.987 BiLSTM-Attention 17.39 5.38 4.19 0.830 0.987 CNN-BiLSTM 21.60 8.70 7.13 0.738 0.981 2 LSTM 19.47 9.82 9.79 0.787 0.985 BiLSTM 19.69 9.27 8.81 0.782 0.985 BiLSTM-Attention 19.23 7.41 6.08 0.792 0.985 CNN-BiLSTM 23.27 11.72 11.24 0.696 0.979 3 LSTM 19.98 8.80 7.80 0.775 0.983 BiLSTM 19.82 8.23 6.80 0.779 0.984 BiLSTM-Attention 20.14 8.04 6.46 0.772 0.983 CNN-BiLSTM 23.15 10.81 9.58 0.699 0.978 27 LSTM 25.40 14.54 13.33 0.635 0.974 BiLSTM 25.92 14.54 12.88 0.620 0.973 BiLSTM-Attention 25.69 13.18 10.92 0.627 0.973 CNN-BiLSTM 26.00 13.75 12.25 0.618 0.972
下载: 导出CSV
表 3 各模型在L&S数据集上多天预测性能比较
Table 3. Comparison of multi-day forecast performance of each model on L&S dataset
预测天数 模型 Metrics RMSE /sfu MAE/sfu MAPE/ (%) R2 R 1 LSTM 14.29 11.26 9.04 0.625 0.993 BiLSTM 14.23 11.02 8.71 0.628 0.993 BiLSTM-Attention 14.75 11.10 8.66 0.600 0.993 CNN-BiLSTM 16.56 13.64 11.45 0.496 0.991 2 LSTM 17.95 13.97 11.30 0.405 0.989 BiLSTM 16.68 12.49 9.92 0.486 0.991 BiLSTM-Attention 16.79 13.04 10.46 0.479 0.990 CNN-BiLSTM 19.23 15.95 13.71 0.316 0.990 3 LSTM 19.14 15.11 12.07 0.325 0.985 BiLSTM 17.43 13.99 11.39 0.440 0.990 BiLSTM-Attention 18.16 14.88 12.19 0.392 0.989 CNN-BiLSTM 20.21 16.40 13.63 0.247 0.987 27 LSTM 24.52 19.84 15.30 –0.065 0.982 BiLSTM 24.11 19.40 15.26 –0.030 0.982 BiLSTM-Attention 24.19 19.14 14.75 –0.037 0.982 CNN-BiLSTM 23.85 19.64 15.87 –0.008 0.982
下载: 导出CSV
-
[1] 马瑞平, 纪巧, 徐寄遥. 太阳F10.7指数准27天振荡的小波分析[J]. 空间科学学报, 2007, 27(2): 89-95 doi: 10.11728/cjss2007.02.089MA Ruiping, JI Qiao, XU Jiyao. Wavelet analysis of quasi-27-day oscillations in the solar index F10.7[J]. Chinese Journal of Space Science, 2007, 27(2): 89-95 doi: 10.11728/cjss2007.02.089 [2] 刘四清, 钟秋珍, 温靖, 等. 太阳10.7 cm射电流量中期预报模型研究(Ⅰ)[J]. 空间科学学报, 2010, 30(1): 1-8 doi: 10.11728/cjss2010.01.001LIU Siqing, ZHONG Qiuzhen, WEN Jing, et al. Modeling research of 10.7 cm solar radio flux 27-day forecast (I)[J]. Chinese Journal of Space Science, 2010, 30(1): 1-8 doi: 10.11728/cjss2010.01.001 [3] 温靖, 钟秋珍, 刘四清. 太阳10.7 cm射电流量中期预报模型研究(Ⅱ)[J]. 空间科学学报, 2010, 30(3): 198-204 doi: 10.11728/cjss2010.03.198WEN Jing, ZHONG Qiuzhen, LIU Siqing. Model research of 10.7cm solar radio flux 27-day forecast (II)[J]. Chinese Journal of Space Science, 2010, 30(3): 198-204 doi: 10.11728/cjss2010.03.198 [4] WANG H B, XIONG J N, ZHAO C Y. The mid-term forecast method of solar radiation index[J]. Chinese Astronomy and Astrophysics, 2015, 39(2): 198-211 doi: 10.1016/j.chinastron.2015.04.010 [5] 高扬, 吕建永, 王明, 等. 基于深度学习的太阳F10.7辐射通量的短期预报研究[J]. 天文学报, 2022, 63(1): 11GAO Yang, LV Jianyong, WANG Ming, et al. Short-term prediction of solar F10.7 radiation flux based on deep learning[J]. Acta Astronomica Sinica, 2022, 63(1): 11 [6] VALDÉS J J, NIKOLIĆ L, TAPPING K. Machine learning approaches for predicting the 10.7 cm radio flux from solar magnetogram data[C]//2019 International Joint Conference on Neural Networks (IJCNN). Budapest, Hungary: IEEE, 2019 [7] HUANG C, LIU D D, WANG J S. Forecast daily indices of solar activity, F10.7, using support vector regression method[J]. Research in Astronomy and Astrophysics, 2009, 9(6): 694 doi: 10.1088/1674-4527/9/6/008 [8] LUO J Q, ZHU H B, JIANG Y, et al. The 10.7-cm radio flux multistep forecasting based on empirical mode decomposition and back propagation neural network[J]. IEEJ Transactions on Electrical and Electronic Engineering, 2020, 15(4): 584-592 doi: 10.1002/tee.23092 [9] ABRAHAM S, ANIYAN A K, KEMBHAVI A K, et al. Detection of bars in galaxies using a deep convolutional neural network[J]. Monthly Notices of the Royal Astronomical Society, 2018, 477(1): 894-903 doi: 10.1093/mnras/sty627 [10] 杨旭, 朱亚光, 杨升高, 等. LSTM神经网络在太阳F10.7射电流量中期预报中的应用[J]. 空间科学学报, 2020, 40(2): 176-185 doi: 10.11728/cjss2020.02.176YANG Xu, ZHU Yaguang, YANG Shenggao, et al. Application of LSTM neural network in F10.7 solar radio flux mid-term forecast[J]. Chinese Journal of Space Science, 2020, 40(2): 176-185 doi: 10.11728/cjss2020.02.176 [11] LUO J Q, ZHU L C, ZHU H B, et al. A new approach for the 10.7-cm solar radio flux forecasting: based on empirical mode decomposition and LSTM[J]. International Journal of Computational Intelligence Systems, 2021, 14(1): 1742-1752 doi: 10.2991/ijcis.d.210602.001 [12] CINAR Y G, MIRISAEE H, GOSWAMI P, et al. Period-aware content attention RNNs for time series forecasting with missing values[J]. Neurocomputing, 2018, 312: 177-186 doi: 10.1016/j.neucom.2018.05.090 [13] LIU H, LIU C, WANG J T L, et al. Predicting solar flares using a long short-term memory network[J]. The Astrophysical Journal, 2019, 877(2): 121 doi: 10.3847/1538-4357/ab1b3c [14] ZHANG W T, ZHAO X H, FENG X S, et al. Predicting the daily 10.7-cm solar radio flux using the long short-term memory method[J]. Universe, 2022, 8(1): 30 doi: 10.3390/universe8010030 [15] LUO J Q, ZHU L C, ZHANG K L, et al. Forecasting the 10.7-cm solar radio flux using deep CNN-LSTM neural networks[J]. Processes, 2022, 10(2): 262 doi: 10.3390/pr10020262 [16] STEVENSON E, RODRIGUEZ-FERNANDEZ V, MINISCI E, et al. A deep learning approach to solar radio flux forecasting[J]. Acta Astronautica, 2022 , 193: 595-606 [17] ZHENG Y F, LI X B, YAN S N, et al. Multiclass solar flare forecasting models with different deep learning algorithms[J]. Monthly Notices of the Royal Astronomical Society, 2023, 521(4): 5384-5399 doi: 10.1093/mnras/stad839 [18] SCHUSTER M, PALIWAL K K. Bidirectional recurrent neural networks[J]. IEEE Transactions on Signal Processing, 1997, 45(11): 2673-2681 doi: 10.1109/78.650093 -
-
下载:
图(10) / 表(3)
计量
- 文章访问数: 670
- HTML全文浏览量: 234
- PDF下载量: 107
-
被引次数:
0(来源:Crossref)
0(来源:其他)
