How well can we forecast local magnetic ground perturbations with existing space weather monitoring resources?

Abstract. In this study we examine how a deep-learning based forecast of local, ground-based geomagnetic field variations trained on solar wind parameters available in real time might be improved by including information contained in an accurate forecast of solar wind conditions. This is accomplished using a long short-term memory (LSTM) model together with magnetic field measurements made at the Rørvik magnetometer station in Mid-Norway. We use Advanced Composition Explorer (ACE) satellite measurements of solar wind and interplanetary magnetic field (IMF) conditions at the first Sun-Earth Lagrange point, and historical lists of coronal mass ejection (CME) impacts at Earth to train and validate the LSTM model. We find that accurate information about the IMF B_z component and solar wind speed are important for obtaining a reasonably accurate (r² ≥ 0.5) forecast of local geomagnetic activity over forecasting horizons beyond ~ 3 h. Information about CME arrival time is only important when simultaneously accompanied by accurate, relatively high-resolution information about IMF B_z. In the absence of the latter, CME arrival time information does not contribute to model performance. This empirical result amounts to a quantitative demonstration of the widely recognized impact of IMF orientation on CME geoeffectiveness. This result also highlights that new innovations, probably in the form of new prediction capabilities of conditions in interplanetary space, will be required to produce accurate forecasts of local geomagnetic disturbances beyond a forecast horizon of 1 h.