基于贝叶斯优化学习模型的电励磁同步电机控制方法研究-学报杂志社

期刊信息

合肥工业大学（自然科学版），2026年5月，第49卷第5期：613-620

DOI: 10.3969/j.issn.1003-5060.2026.05.006

作者信息

秦学丰 $ ^{1} $，李国丽 $ ^{2} $，刘方 $ ^{2} $，李泽华 $ ^{2} $

(1. 埃尔朗根-纽伦堡大学电力电子信息学院，德国埃尔朗根 91058；2. 安徽大学电气工程与自动化学院，安徽合肥 230601)

摘要和关键词

摘要: 电励磁同步电机(electrically excited synchronous motor, EESM)可通过相电流的精确控制来输出期望的输出扭矩和角速度。与离线方法相比,基于机器学习的在线控制方法具有实时性好、准确性高和控制模型复杂度低的优点,但存在参数数量多、设置难的问题。文章提出一种基于贝叶斯优化(Bayesian optimization, BO)学习模型的电励磁同步电机控制方法,采用BO中的熵搜索和期望提升方法对模型自动调参。结果表明:所提方法能够实现学习模型参数的自动调节,提高了学习模型的精确度,有效减少重复试验次数;在大样本量训练下,与高斯过程相比,神经网络精确性更高、时间复杂度更低,但内存占用更多;在小样本量训练下,高斯过程精确性更高、计算成本与神经网络相差不大,内存占用更少。

关键词: 电励磁同步电机；参考电流；机器学习；高斯过程回归（GPR）；神经网络；贝叶斯优化（BO）

Authors

QIN Xuefeng $ ^{1} $, LI Guoli $ ^{2} $, LIU Fang $ ^{2} $, LI Zehua $ ^{2} $

(1. Department of Electrical Engineering, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen 91058, Germany; 2. School of Electrical Engineering and Automation, Anhui University, Hefei 230601, China)

Abstract and Keywords

Abstract: An electrically excited synchronous motor (EESM) can achieve the desired torque and angular velocity by precisely controlling the phase current. Compared to offline methods, machine learning-based online control methods have advantages such as real-time performance, high accuracy, and low model complexity. However, they face challenges related to a large number of parameters and difficulties in setting them. Therefore, this paper proposes an EESM control method based on the Bayesian optimization (BO) learning model. This method utilizes entropy search and expectation enhancement techniques in BO to automatically adjust the model parameters. Simulation analysis results demonstrate that the proposed method can achieve automatic adjustment of the learning model parameters, improve the accuracy of the learning model, and effectively reduce the number of repeated experiments. Furthermore, the comparison results indicate that when the training sample size is large, neural networks outperform Gaussian processes in terms of accuracy and time complexity, but they require more memory. On the other hand, when the training sample size is small, Gaussian processes exhibit higher accuracy, similar computational costs to neural networks, and lower memory usage.

Keywords: electrically excited synchronous motor (EESM); reference current; machine learning; Gaussian process regression (GPR); neural network; Bayesian optimization (BO)

基金信息

国家自然科学基金区域创新联合基金重点资助项目（U23A20647）

合肥工业大学学报(自然科学版)

导航菜单

基于贝叶斯优化学习模型的电励磁同步电机控制方法研究