• 校内登录

个人信息 更多+
  • 教师姓名: 韦玉麒
  • 性别: 男
  • 职称: 教授
  • 博士生导师: 是
  • 硕士生导师: 是
  • 学历: 博士研究生毕业
  • 学位: 博士
  • 所在单位: 电气工程学院
  • 电子邮箱:
  • 办公地点: 兴庆校区:东一楼西140
    创新港校区:3号巨构3-3201

团队新闻

当前位置: 中文主页 - 团队新闻

团队论文被IEEE Transactions on Power Electronics期刊录用

发布时间:2026-11-10
点击次数:
发布时间:
2026-11-10
文章标题:
团队论文被IEEE Transactions on Power Electronics期刊录用
内容:

    团队论文“A Transient Hysteresis Model Based on TCN-BiLSTM and Few-Shot Transfer Learning”被IEEE Transactions on Power Electronics期刊录用(中科院一区TOP期刊)。其中学生王川阳是该论文的第一作者。论文摘要如下:

    Abstract: This study proposes a transient hysteresis model based on TCN-BiLSTM, together with a few-shot transfer learning strategy, using the MagnetX database. The model uses magnetic flux density, its increment, temperature, known magnetic-field values, and a mask sequence as inputs. Sliding-window sampling is used to improve data utilization under three prediction scenarios with different known-sequence lengths. Residual dilated convolutions extract local waveform features and long-range dependencies, while Bi-LSTM layers integrate bidirectional temporal information and directly predict the complete magnetic-field sequence, avoiding autoregressive error accumulation. A physics-informed composite loss jointly constrains sequence and hysteresis-energy errors, while AdamW, cosine annealing and early stopping improve training stability and generalization. For the five Phase 2 materials, the proposed model achieves the 95th percentile error of 7.84% for the H sequence and 1.39% for loss density, placing it among the top-performing teams in the comparison. Transfer learning further improves prediction accuracy under limited data; at 596 KB, it reduces sequence and loss-density errors from 36.04% to 18.94% and from 19.10% to 6.26%, respectively. The results confirm that the proposed method provides accurate transient hysteresis and loss predictions across different materials and offers an efficient solution for rapid modeling of new magnetic materials with limited experimental data.