Spatiotemporal quench prediction of No-Insulation superconducting coils using a GNN-Informer fusion model

Abstract

Abstract Quench prediction is essential for the safe operation of no-insulation high-temperature superconducting coils in fusion magnet systems. This study proposes a spatiotemporal quench prediction model that integrates a Graph Neural Network with Informer to fuse multi-sensor spatial correlations and long-term temporal features. An integrated experimental-simulation dataset was constructed using voltage, temperature, strain, and magnetic-field signals from a no-insulation superconducting coil platform. Results show that the proposed model accurately captures quench evolution and provides a 93 ms earlier warning than the real-time filtered voltage threshold method. The model also maintains robustness under moderate noise and achieves millisecond-level inference. After ONNX-based simplification, quantization, and pruning, the model size is reduced from 16.2 MB to 2.01 MB, with edge-device inference latency below 5 ms. These results demonstrate the potential of the GNN-Informer model for accurate, real-time, and deployable quench early warning in superconducting magnet systems.