【目的】针对模型预测控制权重系数设计困难的问题，本文采用非支配排序遗传算法II(NSGA-II)和贝叶斯优化算法进行权重系数设计。【方法】基于永磁同步电机(PMSM)模型预测转矩控制(MPTC)系统，针对不考虑开关次数控制和考虑开关次数控制两种场景，分别采用NSGA-II和贝叶斯优化算法设计权重系数。不考虑开关次数控制时仅需设计一个权重系数，考虑开关次数控制时需同时设计两个权重系数。基于两种优化算法设计的权重系数，从控制效果、执行时间和内存占用对两种算法进行了对比。【结果】结果表明，对于考虑和不考虑开关次数控制的PMSM MPTC系统，两种权重系数设计算法均可行。NSGA-II得到的使适应度函数值最小的权重系数与贝叶斯优化算法得到的最优权重系数基本相当，控制性能也基本相当，贝叶斯优化算法的控制性能相对略优。【结论】NSGA-II可提供一组适合不同应用场景的Pareto最优解，但其算法复杂度高、计算时间长且占用内存大，适用于动态变化的运行场景。贝叶斯优化算法易于实现、占用资源少，在多控制目标的复杂场景中具有更好的寻优效果和更高的寻优效率。

[Objective] Aiming to the problem of design weighting factor in model predictive control, non-dominated sorting genetic algorithm II (NSGA-II) and Bayesian optimization algorithm are used to design weighting factors in this paper. [Methods] Based on model predictive torque control (MPTC) for permanent magnet synchronous motor (PMSM), NSGA-II and Bayesian optimization algorithm were used to design weighting factors in two scenarios, which were without/with considering switching frequency control, respectively. When without considering switching frequency control, only one weighting factor needed to be designed, when with considering switching frequency control, two weighting factors needed to be designed. And based on the weighting factors designed by two algorithms, a comparison of the two algorithms in terms of control performance, execution time and memory occupancy was carried out. [Results] The results showed that both weighting factors design algorithms were feasible for PMSM MPTC system with/without considering switching frequency control. The weight factors obtained by NSGA-II that minimize the fitness function value were essentially equivalent to the optimal weight factors derived from the Bayesian optimization algorithm, and control performances were basically equivalent. The control performance of Bayesian optimization algorithm was relatively superior. [Conclusion] NSGA-II can provide a set of Pareto optimal solutions suitable for diverse application scenarios. Characterized by high computational complexity, extended processing times, and substantial memory requirements, it is well-suited for dynamically evolving operational environments. And the Bayesian optimization algorithm is easy to implement and doesn’t need too much resource, has better optimization effect and higher optimization efficiency in complex scenarios involving multiple control objectives.

青海省高原汽车电动化与智能化技术重点实验室开放基金(QZDSZ03-202502)