Solar wind is an intermediary in energy transfer from the Sun into the Earth’s magnetosphere, and is considered as a decisive driver of energetic electron dynamics at the geosynchronous orbit (GEO). Based on machine learning technology, several models driven by solar wind parameters have been established to predict GEO electron fluxes. However, the relative contributions of different solar wind parameters on GEO electron fluxes are still unclear. Recently, a feature attribution method, Deep SHapley Additive exPlanations (Deep SHAP) is proposed to open black boxes of machine learning models. In this study, we use the Deep SHAP method to quantify contributions of different solar wind parameters with the artificial neural network (ANN) model proposed by Wang et al. (2023) (https://doi.org/10.1016/j.asr.2022.10.013). Backpropagating the prediction results of this ANN model from 2011 to 2020, SHAP values for four solar wind parameters (IMF Bz, solar wind speed, solar wind dynamic pressure, and proton density) are calculated and comprehensively analyzed. The results suggest that solar wind speed with a lag of 1 day is the most important driver. We further investigate relative roles of different parameters in three specific electron fluxes variation events (corresponding to electron fluxes reaching a local maximum, a local minimum, and unchanged, respectively). The results suggest that high solar wind speed and southward IMF Bz (high dynamic pressures) facilitate increases (decreases) of electron fluxes. These findings help reveal the underlying physical mechanisms of GEO electron dynamics and help develop more accurate prediction models for GEO electron fluxes.

太阳风