Predicting Dual‐Satellite Ionospheric Radial Currents From Single‐Satellite Observations With a BiLSTM Model

Abstract Ionospheric radial currents (IRC) can be derived using two methodologies: (a) the single‐satellite approach: Single‐IRC is deduced from magnetic field data of a single satellite (Swarm‐A); (b) the dual‐satellite approach: Dual‐IRC is calculated by coordinated magnetic field measurements of a pair of satellites (Swarm‐A and ‐C). The magnetic local time (MLT) dependence of IRC is more accurately characterized by the Dual‐IRC than by the Single‐IRC. However, the Single‐IRC can be obtained from multiple missions, whereas the Dual‐IRC is exclusively derived from the Swarm‐A/‐C pair. Consequently, the spatiotemporal coverage of Single‐IRC is broader than that of Dual‐IRC. Therefore, it is necessary to make use of Single‐IRC to obtain more accurate IRCs. To address this, a machine learning model with the bidirectional long short‐term memory (BiLSTM) has been developed. In this supervised learning framework, Dual‐IRC is the target, while the Single‐IRC and 21 relevant parameters are the inputs. Both the BiLSTM predictions (Predict‐IRC) and Dual‐IRC in the testing set are inward/outward at 00–10/10–20 MLT. However, the Single‐IRC peaks outward at 07–12 MLT. The correlation coefficients between Predict‐IRC/Single‐IRC and Dual‐IRC are 0.662/−0.317, respectively. Therefore, the model successfully predicts Dual‐IRC and eliminates the potential errors of Single‐IRC. An analysis of Dual‐IRC, Predict‐IRC, and Single‐IRC under different Kp, solar wind dynamic pressure, and interplanetary magnetic field B z is conducted. At all MLTs, the mean absolute error (MAE) between Predict‐IRC and Dual‐IRC are stable and small, indicating the model's good performance. However, the MAE between Single‐IRC and Dual‐IRC increase under high Kp/intense P d /southward B z .