A Quench Behavior Predictive Model for High Temperature Superconducting Magnet Based on Deep-Learning Neural Network

Abstract

No-insulation (NI) high temperature superconduct-ing (HTS) coils show higher stability than traditionally insulated HTS coils. However, quench remains one of the most crucial issues affecting the safe operation of NI magnets. The quench behaviors in NI coils exhibit inherent complexity since turn-to-turn current redistribution. Low normal zone propagation speed of HTS materials makes it difficult to detect the local hotspot in the early stage of quench, which potentially leads to irreversible damage. In this study, a multi-physical quench behavior predictive model based on Long Short-Term Memory (LSTM) network for HTS NI coils is proposed. Quench data is obtained from an electromagnetic-thermal coupled numerical model with different quench initial locations. By leveraging multi-physical signals as input, the model can predict the dynamic quench behaviors over a future period of time, including temperature, azimuthal current, radial current density and magnetic field. Additionally, the model is capable of predicting quench behaviors at different spatial locations within the coil, achieving a prediction speed of 0.002 seconds and a prediction error below 0.2%. This method demonstrates promise for early quench detection using multi-physical signals and for enabling a timely protection system response.