Knowledge discovery with additive attribution methods for geomagnetic index prediction

Abstract

During a solar cycle, the Sun emits various ejections such as Coronal Mass Ejections (CME) and higher-pressure streams caused by Corotating Interaction Regions (CIR). These ejections induce geomagnetic storms that are known to cause various measurable and often adverse effects. These effects are often modeled with a proxy such as a Geomagnetic Disturbance Index (GMD). One such GMD index is eh, which is derived from the horizontal component of the locally measured ground-based geomagnetic field, and is a proxy index for induced electric field, and a direct driver of Geomagnetically Induced Currents (GIC). Solar physics has a strong reliance on empirical modeling, to which deep learning techniques such as Deep Neural Networks (DNNs) have been applied with moderate success. DNNs have great potential to accurately model these complex relationships due to their high capacity. However, DNNs are black boxes, and the correlations used by these models are nearly impossible to discern. With the advent of additive attribution methods such as DeepSHAP, it is now possible to analyse some of the underlying patterns and relationships used by these models in making their predictions. This study documents the use of a DNN model to predict the eh index at two different latitudes (Abisko, Sweden and Hermanus, South Africa) using solar wind drivers as input and subsequent DeepSHAP attribution. We evaluate the feasibility of such a set of techniques as a knowledge discovery tool to be applied to problems with an abundance of data, but limited knowledge of the effective relationships between input and output drivers of that system. We trained a Multi-layer Perceptron (MLP) with time-shifted inputs to predict the eh index individually at each latitude, simultaneously at both latitudes, and the explicit difference in eh between the latitudes. These models were interpreted using DeepSHAP, and the resulting attributions were compared to the domain literature and known relationships between the solar wind drivers and eh at these latitudes. We found that the DeepSHAP attributions, in the cases where the models accurately predicted eh, consistently met the expectations set by the domain literature.