Scholarlink

<p>Accurate prediction of temperature field of high-speed permanent magnet synchronous motor (PMSM) under extreme conditions is the core challenge to ensure its reliable operation. The traditional lumped parameter thermal network (LPTN) model is difficult to characterize the thermal coupling effect of complex space due to oversimplification of heat conduction path. Although the simple data-driven method can capture the dynamic time series characteristics, it is easy to lead to thermodynamic paradox due to the lack of physical law constraints. In view of the above problems, this paper proposes a hybrid modeling framework that combines graph neural network (GNN), long short-term memory (LSTM) network and physical guidance mechanism. Based on the topology of the motor thermal network, a heterogeneous graph structure is constructed. The cross-scale spatial heat conduction characteristics are extracted by multi-layer graph convolution operation. At the same time, the time series coupling relationship between load current and speed is analyzed by bidirectional LSTM module, and the dynamic mapping of working condition-temperature response is established. In order to strengthen the physical consistency, the heat balance equation is embedded into the loss function to construct a differential physical constraint layer. Experiments show that the mean square error (MSE) of temperature prediction of this model is less than 2℃ under different working conditions, which verifies the good robustness and generalization of the model.</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

Temperature Prediction Method of Permanent Magnet Motor Based on GNN-LSTM and Lumped Parameter Thermal Network

  • 期刊论文

  • 发表日期：

    2025年9月4日

查看全文
  • Pin Lv,
  • Zehua Shang,
  • Ning Wang,
  • Jinming Fu,
  • Chenxi Zhang

摘要

Accurate prediction of temperature field of high-speed permanent magnet synchronous motor (PMSM) under extreme conditions is the core challenge to ensure its reliable operation. The traditional lumped parameter thermal network (LPTN) model is difficult to characterize the thermal coupling effect of complex space due to oversimplification of heat conduction path. Although the simple data-driven method can capture the dynamic time series characteristics, it is easy to lead to thermodynamic paradox due to the lack of physical law constraints. In view of the above problems, this paper proposes a hybrid modeling framework that combines graph neural network (GNN), long short-term memory (LSTM) network and physical guidance mechanism. Based on the topology of the motor thermal network, a heterogeneous graph structure is constructed. The cross-scale spatial heat conduction characteristics are extracted by multi-layer graph convolution operation. At the same time, the time series coupling relationship between load current and speed is analyzed by bidirectional LSTM module, and the dynamic mapping of working condition-temperature response is established. In order to strengthen the physical consistency, the heat balance equation is embedded into the loss function to construct a differential physical constraint layer. Experiments show that the mean square error (MSE) of temperature prediction of this model is less than 2℃ under different working conditions, which verifies the good robustness and generalization of the model.