Scholarlink

<p>Plasma disruption poses a critical challenge to the steady-state operation of tokamaks. Based on 201 discharge shots from the Nanchang Spherical Tokamak (NCST), this study presents a systematic evaluation of disruption prediction models that integrate sliding window statistical features with machine learning. From eight diagnostic signals, local mean and standard deviation within a 5 ms window are extracted to construct a 24-dimensional feature space. Four models—Random Forest, Logistic Regression, Support Vector Machine and LightGBM—are compared under a 20 ms warning window, with hyperparameters optimized via cross-validation and cost-sensitive learning. Results indicate that LightGBM outperforms the others, achieving 96.7% accuracy, 100% recall, and an F1 score of 0.960. Feature importance analysis highlights loop voltage, toroidal magnetic field, and plasma stored energy as the most significant contributors, which aligns with the physical mechanisms of disruptions. This work validates the feasibility of disruption prediction on NCST and provides a model reference for the real-time warning system of the upgraded NCST-U device.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Plasma Disruption Prediction on the Nanchang Spherical Tokamak: A Comparative Study of Models Based on Sliding Window Statistical Features

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Li Zhang,
  • Xincheng Xiong,
  • Xiaochang Chen,
  • Sanqiu Liu,
  • Xiaolan Liu

摘要

Plasma disruption poses a critical challenge to the steady-state operation of tokamaks. Based on 201 discharge shots from the Nanchang Spherical Tokamak (NCST), this study presents a systematic evaluation of disruption prediction models that integrate sliding window statistical features with machine learning. From eight diagnostic signals, local mean and standard deviation within a 5 ms window are extracted to construct a 24-dimensional feature space. Four models—Random Forest, Logistic Regression, Support Vector Machine and LightGBM—are compared under a 20 ms warning window, with hyperparameters optimized via cross-validation and cost-sensitive learning. Results indicate that LightGBM outperforms the others, achieving 96.7% accuracy, 100% recall, and an F1 score of 0.960. Feature importance analysis highlights loop voltage, toroidal magnetic field, and plasma stored energy as the most significant contributors, which aligns with the physical mechanisms of disruptions. This work validates the feasibility of disruption prediction on NCST and provides a model reference for the real-time warning system of the upgraded NCST-U device.