Scholarlink

Aiming at the shortcomings of traditional lightning over-voltage inversion methods, such as low computational efficiency and poor generalization, this paper proposes a new inversion architecture integrating the physics-constrained generative adversarial network (GAN) and the bidirectional long short-term memory (Bi-LSTM) with attention mechanism. To generate high-fidelity simulated data, this method imposes physical constraints on the GAN by embedding the lightning current dual-exponential lightning current model and the nonlinear voltage-current characteristic of arresters. Measured samples are combined with this simulated data to train the Bi-LSTM with attention mechanism. This network leverages its bidirectional one-dimensional convolutional layers to extract nanosecond-level transient features and LSTM layers to capture microsecond-level decay characteristics, establishing a mapping model from arrester tail current to head voltage. Compared to traditional circuit-model-simulation-based inversion methods, the proposed approach achieves an inversion time of only 100 ms, a waveform correlation coefficient of 0.9, and a wavefront time error of less than 3%. This provides a new solution for online over-voltage identification in smart grids.

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

Lightning Parameter Inversion Using Physics Constrained GAN and Bidirectional LSTM

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiangyang Peng,
  • Kang Xie,
  • Yuan Zhou,
  • Kang Wang,
  • Rui Wang,
  • Weimin Guan,
  • Jianbin Fan

摘要

Aiming at the shortcomings of traditional lightning over-voltage inversion methods, such as low computational efficiency and poor generalization, this paper proposes a new inversion architecture integrating the physics-constrained generative adversarial network (GAN) and the bidirectional long short-term memory (Bi-LSTM) with attention mechanism. To generate high-fidelity simulated data, this method imposes physical constraints on the GAN by embedding the lightning current dual-exponential lightning current model and the nonlinear voltage-current characteristic of arresters. Measured samples are combined with this simulated data to train the Bi-LSTM with attention mechanism. This network leverages its bidirectional one-dimensional convolutional layers to extract nanosecond-level transient features and LSTM layers to capture microsecond-level decay characteristics, establishing a mapping model from arrester tail current to head voltage. Compared to traditional circuit-model-simulation-based inversion methods, the proposed approach achieves an inversion time of only 100 ms, a waveform correlation coefficient of 0.9, and a wavefront time error of less than 3%. This provides a new solution for online over-voltage identification in smart grids.