基于输入凸神经网络的IPT系统电压预测控制  

Predictive Control for Output Voltage of IPT System Based on Input Convex Neural Network

作　　者：陈伟华[1] 姜兆迪 CHEN Wei-hua;JIANG Zhao-di(Faculty of Electrical and Control Engineering,Liaoning Technical University,Huludao 125000,China)

机构地区：[1]辽宁工程技术大学电气与控制工程学院,辽宁葫芦岛125000

出　　处：《控制工程》2022年第11期2010-2017,共8页Control Engineering of China

基　　金：2019年辽宁省教育厅科学技术研究创新团队项目(LT2019007);2020年辽宁省教育厅科学研究青年科技人才“育苗”项目(LJ2020QNL019)。

摘　　要：感应式电能传输(IPT)系统的输出电压容易受到负载、互感等系统参数变化的影响。为了提高该系统输出电压的实时调节能力,提出了一种基于双深度输入凸神经网络的模型预测(DDICNN-MPC)控制器。以LCL-LCL结构的IPT系统为例,基于深度输入凸神经网络分别训练了系统状态预测网络、系统误差预测网络,结合二者使用交互递推的方式建立了IPT系统多步误差预测模型,将控制器中的优化问题转化为凸优化问题,然后使用梯度下降算法求解该问题。仿真结果表明,控制器的稳态误差可控制在1%以内,且相比于PID控制器与传统MPC控制器,DDICNN-MPC控制器对互感和负载扰动具有良好的鲁棒性,在谐振参数发生变化时仍有一定的调节能力,为实现感应式电能传输系统输出电压的实时调节提供了一种行之有效的方法。The output voltage of an inductive power transmission(IPT) system is easily affected by changes in system parameters such as load and mutual inductance. In order to improve the real-time regulation capability of the output voltage, a model predictive controller based on dual depth input convex neural network(DDICNN-MPC) is proposed. Taking the LCL-LCL structure of the IPT system as an example, the system state prediction network and the system error prediction network are trained separately based on the deep input convex neural network, and the multi-step error prediction model of the IPT system is established by using interactive recursione method, which transforms the optimization problem in the controller into a convex optimization problem. Then gradient descent algorithm is used to solve this optimization problem. The simulation results show that the steady-state error of the DICCN-MPC is less than 1%. Compared with the PID controller and traditional MPC, DICCN-MPC has good robustness to load disturbance and mutual inductance disturbance. This controller still has a certain regulation ability when the resonance parameter changes. It provides an effective method for the real-time regulation of the output voltage of the IPT system.

关 键 词：输入凸神经网络 模型预测控制 感应式电能传输 电压调节 梯度下降算法 

分 类 号：TM133[电气工程—电工理论与新技术]