Scholarlink

<p>In the design of Permanent Magnet Synchronous Motor (PMSMs), it is essential to simultaneously satisfy various performance requirements such as high output power, high efficiency, and low cogging torque. Finite Element Analysis (FEA) is commonly employed to meet these requirements due to its high accuracy. However, the high computational cost of FEA poses limitations when exploring a large design space. To overcome this limitation, surrogate model-based performance prediction techniques have recently gained attention. In this study, we conduct a comparative analysis of the performance of Kriging and Deep Neural Network (DNN) models for predicting key electromagnetic characteristics of PMSMs. Design variables were sampled using Latin Hypercube Sampling (LHS), and FEA was performed to obtain electromagnetic features such as d-axis and q-axis inductances, permanent magnet flux linkage, and core loss to construct the training dataset. The prediction accuracy, data efficiency, and generalization performance of the two surrogate models were then evaluated. As a result, the DNN exhibited overall higher prediction accuracy, scalability, and generalization performance than the Kriging model. However, when the number of training samples was limited, the Kriging model still delivered competitive performance. The predictive performance of the two surrogate models was evaluated using an independent test dataset.</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

Comparative study of kriging and deep neural networks as surrogate models for parameter prediction of PMSM

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Yong-Min Lee,
  • Dong-Hoon Ko,
  • Min-Ro Park

摘要

In the design of Permanent Magnet Synchronous Motor (PMSMs), it is essential to simultaneously satisfy various performance requirements such as high output power, high efficiency, and low cogging torque. Finite Element Analysis (FEA) is commonly employed to meet these requirements due to its high accuracy. However, the high computational cost of FEA poses limitations when exploring a large design space. To overcome this limitation, surrogate model-based performance prediction techniques have recently gained attention. In this study, we conduct a comparative analysis of the performance of Kriging and Deep Neural Network (DNN) models for predicting key electromagnetic characteristics of PMSMs. Design variables were sampled using Latin Hypercube Sampling (LHS), and FEA was performed to obtain electromagnetic features such as d-axis and q-axis inductances, permanent magnet flux linkage, and core loss to construct the training dataset. The prediction accuracy, data efficiency, and generalization performance of the two surrogate models were then evaluated. As a result, the DNN exhibited overall higher prediction accuracy, scalability, and generalization performance than the Kriging model. However, when the number of training samples was limited, the Kriging model still delivered competitive performance. The predictive performance of the two surrogate models was evaluated using an independent test dataset.