A Deep Learning‐Based Approach to Forecast the Onset of Magnetic Substorms

Abstract

The auroral substorm has been extensively studied over the last six decades. However, our understanding of its driving mechanisms is still limited and so is our ability to accurately forecast its onset. In this study, we present the first deep learning‐based approach to predict the onset of a magnetic substorm, defined as the signature of the auroral electrojets in ground magnetometer measurements. Specifically, we use a time history of solar wind speed (Vx), proton number density, and interplanetary magnetic field (IMF) components as inputs to forecast the occurrence probability of an onset over the next 1 hr. The model has been trained and tested on a data set derived from the SuperMAG list of magnetic substorm onsets and can correctly identify substorms ∼75% of the time. In contrast, an earlier prediction algorithm correctly identifies ∼21% of the substorms in the same data set. Our model's ability to forecast substorm onsets based on solar wind and IMF inputs prior to the actual onset time, and the trend observed in IMF Bz prior to onset together suggest that a majority of the substorms may not be externally triggered by northward turnings of IMF. Furthermore, we find that IMF Bz and Vx have the most significant influence on model performance. Finally, principal component analysis shows a significant degree of overlap in the solar wind and IMF parameters prior to both substorm and nonsubstorm intervals, suggesting that solar wind and IMF alone may not be sufficient to forecast all substorms, and preconditioning of the magnetotail may be an important factor.

A Deep Learning‐Based Approach for Modeling the Dynamics of AMPERE Birkeland Currents

“…In particular, we aim to capture two features as accurately as possible : (1) the reconfiguration of the FAC system from one steady state to another and (2) the response of FACs to a change in IMF at different locations. Both these tasks require a model that is suitable for capturing complex features in a temporal data set, and a ResNet CNN is one such model (Fawaz et al., 2019; He, Zhang, et al., 2016; Maimaiti et al., 2019). Furthermore, the ResNet CNN architecture is less prone to the adverse effects associated with deeper architectures and is relatively easier to optimize (Glorot & Bengio, 2010).…”

“…Finally, we split the data set into three parts: train (90%), validation (9%), and test (1%) in chronological order, similar to the approach taken in Maimaiti et al. (2019). Such a split prevents the model from overfitting as the test and validation time periods are completely independent of the training time period.…”

“…We chose the input parameters based on prior knowledge (e.g., Gjerloev, 2012; Iyemori, 1990; Laundal et al., 2016; Laundal, Finlay, et al., 2018; Weimer et al., 2003) of their role in influencing FACs as well as their availability over extended periods of time. However, we didn't use composite solar wind‐magnetosphere coupling functions such as d Φ/ dt (Newell et al., 2007) as inputs to our model since previous studies have shown that deep learning models can replicate the behavior of such functions and their addition doesn't substantially improve the model's performance any further (e.g., Maimaiti et al., 2019).…”

“…In this study, we take advantage of the recent advancements in deep learning to model the time‐dependent evolution of AMPERE FACs in the Northern Hemisphere. The large and continuously available database of global Birkeland current patterns provided by the AMPERE project is ideally suitable for training deep learning architectures that are capable of capturing such complicated behavior (LeCun et al., 2015; Maimaiti et al., 2019). To this end, we use a 60‐min time history of solar wind bulk speed (Vx), proton number density (Np), IMF components (Bx, By, and Bz), and geomagnetic and solar indices (Sym‐H, Asym‐H, SuperMAG AL/AU, and F10.7 flux) along with the month (indicative of season) as inputs to a convolutional neural network (CNN) model to estimate global Birkeland currents at a given moment.…”

Deep Neural Networks With Convolutional and LSTM Layers for SYM‐H and ASY‐H Forecasting

“…Deep Learning (DL) and ANNs have been used as well to forecast the onset of magnetic storms, as presented in the work of Maimaiti et al. (2019). They use a time history of solar wind speed, proton density, and IMF as inputs to forecast the occurrence probability of the onset of magnetic substorms over the next hour.…”