Forecasting the Dst Index with Temporal Convolutional Network and Integrated Gradients

• Published in: Solar physics Vol. 299; no. 7; p. 98
• Main Authors: Liu, Junyan, Shen, Chenglong, Wang, Yang, Xu, Mengjiao, Chi, Yutian, Zhong, Zhihui, Mao, Dongwei, Zhang, Zhiyong, Wang, Can, Liu, Jiajia, Wang, Yuming
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01-07-2024; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Astrophysics and Astroparticles; Atmospheric Sciences; DST Index; Electric fields; Errors; Forecast accuracy; Forecasting models; Geomagnetic disturbances; Geomagnetism; Interplanetary electric fields; Machine learning; Magnetic fields; Nonlinear systems; Observation times; Parameters; Physical properties; Physics; Physics and Astronomy; Prediction models; Proton density (concentration); Protons; Solar magnetic field; Solar wind; Space Exploration and Astronautics; Space Sciences (including Extraterrestrial Physics; Space weather; Weather forecasting; Wind speed; Wind velocities; Space-weather forecast; Machine learning; Geomagnetic storm
• ISSN: 0038-0938; 1573-093X
• DOI: 10.1007/s11207-024-02340-9

The Disturbance Storm Time (Dst) Index stands as a crucial geomagnetic metric, serving to quantify the intensity of geomagnetic disturbances. The accurate prediction of the Dst index plays a pivotal role in mitigating the detrimental effects caused by severe space-weather events. Therefore, Dst prediction has been a long-standing focal point within the realms of space physics and space-weather forecasting. In this study, a Temporal Convolutional Network (TCN) is deployed in tandem with the Integrated Gradient (IG) algorithm to predict the Dst index and scrutinize its associated physical processes. With these two components, our model can give the contribution of each input parameter to the outcome along with the forecast. The TCN component of our model utilizes interplanetary observational data, encompassing the vector magnetic field, solar-wind velocity, proton temperature, proton density, interplanetary electric field, and other relevant parameters for forecasting Dst indices. Despite the disparity in test sets, our model’s forecast accuracy approximates the error levels of the prior models. Remarkably, the prediction error of these machine-learning models has become comparable to the inherent error between the Dst index itself and the actual ring-current strength.
To understand the physical process behind the forecasting model, the IG algorithm was applied in our prediction model, in an attempt to analyze the underlying physical process of the machine-learning black box. In the temporal dimension, it is evident that the more recent the time, the more substantial the influence on the final prediction. Regarding the physical parameters, besides the historical Dst index itself, the flow pressure, the
z
-component of the magnetic field, and the proton density all significantly contribute to the final prediction. Additionally, IG attributions were analyzed for subsets of data, including different Dst-index ranges, different observation times, and different interplanetary structures. Most of the subsets exhibit an IG matrix with deviations from the mean distribution, which indicates a complex nonlinear system and sensitivity of the prediction to input values. These analyses align with physical reasoning and are in good agreement with previous research. The results affirm that the TCN+IG technique not only enhances space-weather forecast accuracy but also advances our comprehension of the underlying physical processes in space weather.