# A hybrid statistical–machine learning framework for evaluating geomagnetic storm effects on MisrSat2 satellite power subsystems

## TL;DR

This paper presents a new hybrid framework combining statistical and machine learning methods to assess the impact of a geomagnetic storm on a satellite's power system.

## Contribution

The novelty is the integration of physics-based modeling, robust statistics, and interpretable machine learning for scalable space mission diagnostics.

## Key findings

- Solar arrays showed minor current deviations during the storm, with deviations under 4% and within design tolerances.
- The MoE machine learning model outperformed baselines with R2=0.921 and MAE=0.063 A in predicting power subsystem behavior.
- Radiation degradation modeling predicted only 0.32% cumulative loss, consistent with telemetry data showing no radiation-driven anomalies.

## Abstract

This study introduces a hybrid statistical–machine learning framework to evaluate the impact of the May 2024 geomagnetic storm on the power subsystem of the MisrSat-2 satellite. The proposed framework integrates a multi-tiered statistical approach, employing CUSUM for change point detection, z-score for outlier identification, and event-based analysis, with robust validation through Welch’s t-tests, bootstrapping, and Benjamini–Hochberg false discovery rate (BH-FDR) control. On 10 May, near the storm’s onset, the solar arrays showed modest current deviations, after validation, 13 events solar panel-1 and 17 events solar panel-2 were retained, with the largest cluster between 07:05 and 09:25 UTC. whereas the battery subsystem remained stable and buffered fluctuations, maintaining bus integrity. Event-based analysis confirmed that all deviations were small (< 4%) and within design tolerances. Radiation degradation modeling with EQUFLUX predicted only 0.32% cumulative loss for May 2024, align with the absence of measurable radiation-driven signatures in telemetry. Extending beyond descriptive detection, a Mixture of Experts (MoE) machine learning framework achieved superior predictive accuracy (R2 = 0.921, MAE = 0.063 A) compared to baseline models providing interpretable validation of statistical findings. The novelty of this research lies in its integrative approach, merging physics-based modelling, robust statistical methods, and interpretable machine learning to provide a scalable framework for anomaly diagnostics and mission assurance in dynamic space environments.