23 October 2025 in IEEE Access
Predicting high-energy proton flux is essential for radiation-effect protection of satellite devices. We introduce a dual-prediction framework based on Graph Neural Networks (GNN) to model proton flux from both non-time-series and time-series perspectives. In the non-time-series approach, the model directly predicts proton flux at specific times using static inputs, allowing rapid assessment of each feature’s impact on prediction accuracy. In the time-series approach, the model utilizes historical data to capture the temporal dynamics of proton flux. The GNN effectively captures complex spatial correlations among input features, particularly under varying solar activity and magnetic-field conditions. Experimental results demonstrate that the GNN performs well in both approaches, with the time-series method achieving higher accuracy for long-term predictions. In both frameworks, the GNN serves as the core modeling tool, structuring spatial locations and geomagnetic variables into a graph to capture dependencies between different physical variables, particularly the complex interactions between solar wind and the geomagnetic field. This graph structure effectively enhances the model’s spatial dependency modeling capacity, aiding in accurately understanding and predicting the distribution patterns of high-energy proton flux.