Search

Abstract In this study, we develop a robust real‐time forecast system for the SYM‐H index using Deep Neural Networks and real‐time Solar Wind measurements along with Interplanetary Magnetic Field parameters. This system provides not only one‐off forecasts but also quantile‐based confidence intervals, offering a range within which the observed SYM‐H values are expected to fall, enhancing forecast reliability and usability. The model is tested both on historical level 2 science‐ready data and on preliminary observations, which are closer to the operational environment that the model is expected to work on, demonstrating the model's robustness and practical utility in real‐time scenarios. The integration of quantile forecasts into SYM‐H prediction models represents a significant advancement, providing decision‐makers with more accurate and trustworthy information to manage the hazard of geomagnetic storms.

Abstract Abduallah et al. (2024b, https://doi.org/10.1029/2023sw003824) proposed a novel approach using a deep neural network model, which includes a graph neural network and a bidirectional LSTM layer, named SYMHnet, to forecast the SYM‐H index one and 2 hr in advance. Additionally, the network also provides an uncertainty quantification of the predictions. While the approach is innovative, there are some areas where the model's design and implementation may not align with best practices in uncertainty quantification and predictive modeling. We focus on discrepancies in the input and output of the model, which can limit the applicability in real‐world forecasting scenarios. This comment aims to clarify these issues, offering detailed insights into how such discrepancies could compromise the model's interpretability and reliability, thereby contributing to the advancement of predictive modeling in space weather research.

Abstract The use of Deep Learning models to forecast geomagnetic storms is achieving great results. However, the evaluation of these models is mainly supported on generic regression metrics (such as the Root Mean Squared Error or the Coefficient of Determination), which are not able to properly capture the specific particularities of geomagnetic storms forecasting. Particularly, they do not provide insights during the high activity periods. To overcome this issue, we introduce the Binned Forecasting Error to provide a more accurate assessment of models' performance across the different intensity levels of a geomagnetic storm. This metric facilitates a robust comparison of different forecasting models, presenting a true representation of a model's predictive capabilities while being resilient to different storms duration. In this direction, for enabling fair comparison among models, it is important to standardize the sets of geomagnetic storms for model training, validation and testing. To do this, we have started from the current sets used in the literature for forecasting the SYM‐H, enriching them with newer storms not considered previously, focusing not only on disturbances caused by Coronal Mass Ejections but also addressing High‐Speed Streams. To operationalize the evaluation framework, a comparative study is conducted between a baseline neural network model and a persistence model, showcasing the effectiveness of the new metric in evaluating forecasting performance during intense geomagnetic storms. Finally, we propose the use of preliminary measurements from ACE to evaluate the model performance in settings closer to an operational real‐time scenario, where the forecasting models are expected to operate.

Abstract In this work, we present an Artificial Neural Network for operational forecasting of the SYM‐H geomagnetic index up to 2 hr ahead using the Interplanetary Magnetic Field, the solar wind plasma features and previous SYM‐H values. Former works that forecast the SYM‐H index use data measured by ACE, in particular from the MAG and SWEPAM instruments. However, the plasma data present a high amount of missing samples. This issue has been addressed in the literature, often using linear interpolation, which leads to a non‐accurate data reconstruction, specially when the features are missing during the most intense periods of a geomagnetic storm. To overcome that issue, we use ACE's Solar Wind Ion Composition Spectrometer (SWICS) data to fill the missing plasma features. To validate this technique, we compare the results of our forecasting model trained using plasma features in two ways: only using SWEPAM and performing linear interpolation and using SWICS to fill the missing values in SWEPAM. Then, both models are evaluated in an operational scenario, when only SWEPAM data are available and interpolation can only be performed if the missing values are surrounded by valid measurements. In both cases, our model outperforms the current literature forecasting the SYM‐H one and 2 hr ahead, yielding the best results when the training has been done using the data completed using SWICS measurements.

Abstract Geomagnetic indices quantify the disturbance caused by the solar activity on a planetary scale or in particular regions of the Earth. Among them, the SYM‐H and ASY‐H indices represent the (longitudinally) symmetric and asymmetric geomagnetic disturbance of the horizontal component of the magnetic field at midlatitude with a 1‐min resolution. Their resolution, along with their relation to the solar wind parameters, makes the forecasting of the geomagnetic indices a problem that can be addressed through the use of Deep Learning, particularly using Deep Neural Networks (DNNs). In this work, we present two DNNs developed to forecast respectively the SYM‐H and ASY‐H indices. Both networks have been trained using the Interplanetary Magnetic Field (IMF) and the related index for the solar storms occurred in the last two solar cycles. As a result, the networks are able to accurately forecast the indices 2 h in advance, considering the IMF and indices values for the previous 200 min. The evaluation of both networks reveals a great forecasting precision, including good predictions for large storms that occurred during the solar cycle 23 and comparing with the persistence model for the period 2013–2020.