Performance evaluation of Neural Network Modelfor different geomagnetic indices forecastingin Solar Cycle 25

Abstract

Geomagnetic storms, characterized by sudden disturbances in Earth’s magnetic field, pose significant risks to technological infrastructure and human activities. This study evaluates the performance of neural network (NN) approaches for predicting different geomagnetic indices during Solar Cycle 25. Nonlinear Autoregressive Networks with Exogenous Inputs (NARX) were trained and tested using solar wind parameters as predictors and geomagnetic indices (Dst, Kp, and Ap)as outputs. The model performed one‑step‑ahead forecasting (1‑day horizon), predicting next‑day geomagnetic indices Dst, Kp, and Ap using daily solar wind parameters from OMNI data (2020‑2025) were evaluated for prediction accuracy, robustness, and computational efficiency using metrics such as root mean square error (RMSE), mean absolute error (MAE), and the cross‑correlation coefficient (R). The results demonstrated strong forecasting capability, achieving Root Mean Squared Error (RMSE) values as low as 0.011 and correlation coefficients up to 0.99 for Dst index predictions.

These results highlight the NARX model’s robustness and accuracy in capturing the complex dynamics of geomagnetic storms. This comprehensive evaluation supports the model’s utility for operational space weather forecasting, providing significant improvements over baseline forecasting methods.