Scholarlink

<p>In dynamic simulations of gas-turbine gas-path performance, complex operating conditions often introduce systematic bias that limits accurate modeling. To address this, a high-precision simulation method is proposed based on error correction using a long short-term memory–kolmogorov–arnold network (LSTM–KAN). A physics-based thermodynamic model was first constructed and validated against field measurements, revealing notable discrepancies in temperature, pressure, and rotational speed. A two-stage LSTM–KAN network was then developed, leveraging LSTM’s strength in temporal modeling and KAN’s nonlinear regression to produce an interpretable compensation term enhancing the model structure. Trained and validated on real operational data, the hybrid model reduces RMSE and MAPE by about 25 % and 30 % on average compared with a standalone LSTM on the test set, and by about 48 % and 52 % on an independent validation dataset, while also outperforming a standalone KAN across all key parameters. Overall, the method significantly improves accuracy and offers a practical foundation for online health monitoring and fault prognosis.</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

High-precision simulation method for gas turbine gas-path performance based on LSTM–KAN network error correction

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Pengzhan Wang,
  • Kun Feng,
  • Peng Zhang

摘要

In dynamic simulations of gas-turbine gas-path performance, complex operating conditions often introduce systematic bias that limits accurate modeling. To address this, a high-precision simulation method is proposed based on error correction using a long short-term memory–kolmogorov–arnold network (LSTM–KAN). A physics-based thermodynamic model was first constructed and validated against field measurements, revealing notable discrepancies in temperature, pressure, and rotational speed. A two-stage LSTM–KAN network was then developed, leveraging LSTM’s strength in temporal modeling and KAN’s nonlinear regression to produce an interpretable compensation term enhancing the model structure. Trained and validated on real operational data, the hybrid model reduces RMSE and MAPE by about 25 % and 30 % on average compared with a standalone LSTM on the test set, and by about 48 % and 52 % on an independent validation dataset, while also outperforming a standalone KAN across all key parameters. Overall, the method significantly improves accuracy and offers a practical foundation for online health monitoring and fault prognosis.