A Regional Ionospheric Storm Forecasting Method Using a Deep Learning Algorithm: LSTM

An ionospheric storm forecasting method was proposed using a deep learning algorithm, LSTM (long short‐term memory). We used the perturbation index to denote the level of an ionospheric storm, deduced from foF2 data, and helped to remove most of the local time and seasonal variations in the ionosphere. In constructing the model, a number of correlated factors were used as inputs, including the properties of coronal mass ejections, solar flare bursts, interplanetary conditions, and geomagnetic and ionospheric states, and the output was whether an ionospheric storm occurred locally in the next 24 hr. Data sets from 2007 to 2014 were used to train the model, and those from 2015 to 2016 were used for validation. The results showed that the model behaved well in most events. The mean precision rate, recall rate, accuracy, and F1 score of the model were 71.7%, 59.7%, 92.7%, and 65.0% in northern China and 78.9%, 56.3%, 96.3% and 65.0% in southern China, respectively. The LSTM forecasting model performed better than other models such as persistence, multiple‐layer perceptron and support vector machine models. Case studies also showed good performance during geomagnetic storms of different strengths. We believe that this model can be beneficial for functional ionospheric storm operation. Plain Language Summary An ionospheric storm forecasting method was proposed using a deep learning algorithm, namely, LSTM (long short‐term memory). The model was trained using an ionospheric measurement data set in China. It was shown that the forecasting model performed better than the persistence, multiple‐layer perceptron and support vector machine models. Key Points A deep learning method is used to build a forecasting model for ionospheric storm A normalized index perturbation index is used to denote if there occur an ionospheric storm or not, which is deduced from the ionospheric foF2 data The long short‐term memory model has a better performance than other models such as persistence, support vector machine and multiple‐layer perceptron model