10:45 AM - 12:15 PM
[PEM16-P07] A Study on Long-Term Prediction of Dst index using Time series deep neural network
Keywords:Geomagnetic storm, Dst index, Machine learning, LSTM, Transformer
In this presentation, we present a prediction model for the Dst index, which indicates the scale of geomagnetic storms, using time series deep learning, a discussion on the predicted values, and future issues.
In recent years, the use of space by the private sector, especially in the low earth orbit region, has been accelerating. Low earth orbit satellites can be developed at a lower cost than geostationary orbit satellites, but they are more susceptible to the space environment because of the need for smaller size and lighter weight. Specifically, software and hardware errors due to space radiation and aerodynamic drag due to the expansion of the upper atmosphere can lead to the loss of the satellite in some cases. These are mainly caused by geomagnetic storms and solar flares. Therefore, long-term prediction of the Dst index, which indicates the magnitude of geomagnetic storms, from solar wind observation data is important to prevent unexpected satellite losses and increased remanufacturing and launch costs.
Gruet et al. proposed a method to predict the Dst index 3 and 6 hours ahead [Gruet et al., 2018]. However, as far as we could find, we could not find any previous studies that applied the Transformer architecture [Vaswani et al., 2017], which has achieved SOTA in various machine learning tasks in the recent past, to predict the Dst index. Therefore, in this study, we used a Transformer-based model in addition to the LSTM-based model proposed by Gruet et al. to compare the predicted values of the Dst index 3 and 6 hours ahead, respectively.
The time series deep learning models used in this experiment were two regression models using LSTM and Transformer. The training data were hourly averaged solar wind observations from August 2001 to October 2022 using the solar wind data available at the NOAA Space Weather Prediction Center. The Dst index, which serves as label data, was obtained from the homepage of the World Data Center for Geomagnetism, Kyoto. RMSE was used as the evaluation index to compare the results with those of previous studies.
The experimental results showed that the model using LSTM had the highest accuracy with RMSE: 6.632 for 3-hour-ahead prediction and the model using Transformer had the highest accuracy with RMSE: 8.669 for 6-hour-ahead prediction, exceeding the index indicated by the previous study. A future problem is the time lag that occurs between the label data and the predictions. It has been pointed out in previous studies that the Dst index is predicted with more delay than actual [Laperre et al., 2020], and the results using the Transformer architecture have not solve this problem and we will discuss methods to solve it.
In recent years, the use of space by the private sector, especially in the low earth orbit region, has been accelerating. Low earth orbit satellites can be developed at a lower cost than geostationary orbit satellites, but they are more susceptible to the space environment because of the need for smaller size and lighter weight. Specifically, software and hardware errors due to space radiation and aerodynamic drag due to the expansion of the upper atmosphere can lead to the loss of the satellite in some cases. These are mainly caused by geomagnetic storms and solar flares. Therefore, long-term prediction of the Dst index, which indicates the magnitude of geomagnetic storms, from solar wind observation data is important to prevent unexpected satellite losses and increased remanufacturing and launch costs.
Gruet et al. proposed a method to predict the Dst index 3 and 6 hours ahead [Gruet et al., 2018]. However, as far as we could find, we could not find any previous studies that applied the Transformer architecture [Vaswani et al., 2017], which has achieved SOTA in various machine learning tasks in the recent past, to predict the Dst index. Therefore, in this study, we used a Transformer-based model in addition to the LSTM-based model proposed by Gruet et al. to compare the predicted values of the Dst index 3 and 6 hours ahead, respectively.
The time series deep learning models used in this experiment were two regression models using LSTM and Transformer. The training data were hourly averaged solar wind observations from August 2001 to October 2022 using the solar wind data available at the NOAA Space Weather Prediction Center. The Dst index, which serves as label data, was obtained from the homepage of the World Data Center for Geomagnetism, Kyoto. RMSE was used as the evaluation index to compare the results with those of previous studies.
The experimental results showed that the model using LSTM had the highest accuracy with RMSE: 6.632 for 3-hour-ahead prediction and the model using Transformer had the highest accuracy with RMSE: 8.669 for 6-hour-ahead prediction, exceeding the index indicated by the previous study. A future problem is the time lag that occurs between the label data and the predictions. It has been pointed out in previous studies that the Dst index is predicted with more delay than actual [Laperre et al., 2020], and the results using the Transformer architecture have not solve this problem and we will discuss methods to solve it.