| Summary: | Machine learning has attracted significant attention for predicting an upcoming geomagnetic storm disturbance using the disturbance storm time (Dst) index. However, studies on machine learning techniques for predicting geomagnetic disturbances have mostly focused on the next step ahead of performance while lacking a comprehensive model explanation between the insights of contributing features and prediction performances have been limited. This study aimed to propose a new framework to include explainable artificial intelligence (AI) by producing an interpretable post-hoc assessment considering multivariate inputs of solar wind parameters, namely interplanetary magnetic field (BzGSM), plasma flow speed (Vsw), proton temperature (Tsw), proton density (N sw) and flow pressure (Pd) which applicable for prediction modelling for the Dst index throughout the 15 years. Varying training algorithms to examine the prediction outcome with Levenberg-Marquardt (LM), Bayesian regularisation (BR), and scaled conj ugate gradient (SCG), were evaluated using the performances of errors analysis including root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and mean square error (MSE). The results revealed the superiority of the NARX model, with the LM training algorithm performing the best with the least errors, thus demonstrating the capabilities to explain the relevant feature insights underlying the AI-predicted outcome at different strengths of impending geomagnetic storms. © 2024 IEEE.
|
|---|