面向动态优化的不规则区域神经网络构建方法  

Construction method of irregular regional neural network for dynamic optimization

作　　者：张琪 吴义忠[1] 邢涛 乔平 ZHANG Qi;WU Yizhong;XING Tao;QIAO Ping(School of Mechanical Science and Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;Beijing Institute of Spacecraft System Engineering,Beijing 100094,China;School of Mechanical Engineering,Suzhou University of Science and Technology,Suzhou 215009,Jiangsu China)

机构地区：[1]华中科技大学机械科学与工程学院,湖北武汉430074 [2]北京空间飞行器总体设计部,北京100094 [3]苏州科技大学机械工程学院,江苏苏州215009

出　　处：《华中科技大学学报（自然科学版）》2023年第12期1-7,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家重点研发项目(2018YFB1700905)。

摘　　要：为实现基于代理模型方法高效求解状态方程隐式表达的动态优化问题,提出一种基于高斯混合模型聚类的空间划分策略,构建状态方程右端式的神经网络模型.根据状态方程右端式的非线性程度,将状态空间划分为多个不规则的局部区域,通过在各个局部区域上分别构建神经网络模型,以获得状态方程右端式的全局神经网络模型,进而基于全局神经网络模型求解动态优化问题.两个工程算例中,基于所提方法得到最优目标函数的近似值与精确值的误差分别为0.0134%和0.0030%,表明该方法保证了基于代理模型方法求解动态优化问题的精度.To efficiently solve the dynamic optimization problem with implicit state equation using the surrogate model-based method,a space partition strategy based on Gaussian mixture model clustering algorithm was proposed to construct the neural network model of the implicit state equation.The state space was divided into several irregular local regions according to the nonlinearity of the right-hand-side functions in the state equation,the global neural network model of the state equation was generated by combining the neural network models of the local regions,then the dynamic optimization problem was solved based on the global neural network model.In the two engineering examples,the errors between the approximate optimal objective function values based on the proposed method and exact optimal objective function values are 0.0134% and 0.0030%,respectively,which show that the proposed method guarantees the solution accuracy of dynamic optimization problem based on the surrogate modelbased method.

关 键 词：动态优化 空间划分 不规则区域 高斯混合模型聚类 神经网络模型 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]