INTERPRETABLE N-BEATS DEEP NETWORKS OF MULTISTEP FORECASTING FOR THE GROUND-BASED GEOMAGNETIC DST INDEX

The importance of geomagnetic disturbances represented by ground earth activities based on the disturbance storm time (Dst index) entails an early forecast of geomagnetic storm occurrence, which could potentially disrupt the system operations. Often, the forecast outcome serves as an essential indicator for operational users who not only require early forecasting prior to incoming geomagnetic storms but also intend to obtain explainable insight and understanding of the generated forecast results. Therefore, a new model architecture, namely neural-basis expansion analysis for interpretable time series (N-BEATS), is proposed that incorporates a more transparent architecture of the deep learning model into producing the multiple steps ahead forecasting of the Dst index. Extensive comparisons among several deep learning models, namely long short-term memory (LSTM), gated recurrent units (GRU), and bidirectional GRU (Bi-GRU) network architectures, will be assessed, considering the model performances, and the impact of forecast variability will be discussed. The superiority of N-BEATS overcomes the state-of-the-art LSTM forecast model in terms of computational resources, and the effectiveness of learning the data of the Dst index pattern could be observed.