Simultaneous Multi–Horizon Forecasting of Dst and Hp60 Geomagnetic Indices via a Multitask LSTM Framework

Authors:
Open Access
Publisher website
Google Scholar
Abstract
This study presents a novel multitask deep learning framework that integrates models designed for simultaneous, multi‐horizon forecasting of Hp60 and Disturbance storm time (Dst) indices with lead times up to 3 hr. Using a Long Short‐Term Memory architecture, the models capture both mid‐latitude variability and the ring current dynamics. Input features include interplanetary magnetic field (IMF) parameters, solar wind plasma data, and historical geomagnetic indices. Model interpretability is addressed through permutation‐based feature importance analysis, providing physical context to the machine learning predictions. The models demonstrate strong predictive performance across all forecast horizons. For the Dst index, they achieve a Root Mean Square Error (RMSE) of 3.81 nT and Pearson correlation coefficient of 0.98 at 1‐hr horizon ( t + 1 hr), remaining robust (RMSE = 9.01 nT, = 0.91) at 3‐hr horizon ( t + 3 hr). More importantly, the models perform well under both quiet and disturbed geomagnetic conditions. However, while they accurately capture storm onset and recovery phases, they underestimate extreme storm intensities. Hp60 predictability is primarily associated with related mid‐latitude indices, such as planetary K (Kp) and ap, that are derived from similar observatory networks and respond to the same underlying geomagnetic disturbances. Dst forecasting is largely influenced by its own historical values (persistence). Southward IMF is also significantly influential for forecasting both indices during disturbed periods, showing that the models learn physically consistent storm‐time behavior. The multitask approach offers a scalable and efficient pathway for future space weather models by leveraging shared representations of common solar wind drivers.

This publication has 38 references indexed in Scilit: