Description:
Extreme space weather events can have large impacts on ground-based infrastructure important to technology-based societies. Machine learning techniques based on interplanetary observations have proven successful as a tool for forecasting global geomagnetic indices, however, few studies have examined local ground magnetic field perturbations. Nowcast and forecast models which predict the magnitude of the horizontal geomagnetic field, |BH|, and its time derivative, dBHdt, at ground level would be valuable for assessing the potential space weather hazard. We attempt to predict the variation of the magnetic field at the three United Kingdom observatories (Eskdalemuir, Hartland and Lerwick) driven by L1 solar wind parameters. The horizontal magnetic field component and its time derivative are predicted from solar wind plasma and interplanetary magnetic field observations using Long Short-Term Memory (LSTM) networks and hybrid Convolutional Neural Network-LSTM models. A 5-fold grid search cross-validation is used for tuning the hyperparameters in each model. Forecasts were made with 5, 15 and 30-min lead times. Models were trained and validated with geomagnetic storm-only data from 1997 to 2016; their outputs were evaluated with the 7–9th September 2017 storms. The forecast models are only able to predict the directly driven parts of geomagnetic storms (not the substorms) and LSTM models generally perform best. We find the |BH| 15- and 30-min forecasts at Lerwick and Eskdalemuir have some predictive power. The 5-min |BH| forecast as well as all the dBHdt models for Eskdalemuir and all the Hartland models were found to have little or no predictive power. This suggests that the machine learning models have better forecasting power at higher latitude (closer to the auroral zones), where the ground magnetic variation field is larger and during storm onset, which is directly driven by changes in the solar wind.