基于分数阶滤波器和BiGRU神经网络的Wiener非线性系统建模与辨识

Modeling and identification of Wiener nonlinear system based on fractional order filter and BiGRU neural network

中文关键词  非线性Wiener系统  分数阶滤波器  双向GRU神经网络  参数辨识  永磁同步电机

英文关键词  nonlinear Wiener system  fractional order filters  bidirectional GRU neural networks  parameter identification  permanent magnet synchronous motor

中文摘要

      本文提出了一种基于分数阶滤波器和双向门控循环单元(BiGRU)神经网络的Wiener非线性系统建模与辨 识方法, 并应用于永磁同步电机的电流和电压预测. 首先, 通过Grunwald-Letnikov ¨ 方法计算分数阶滤波器的未知系 数, 并利用该滤波器对输入数据进行滤波, 去除高频噪声, 增强数据的稳健性; 其次, 为了提升对序列深层次特征的 捕捉能力, 使用BiGRU神经网络同时获取序列数据的过去和未来信息, 并通过自适应动量估计技术更新 BiGRU网 络的参数. 仿真结果表明, 提出的Wiener系统能够有效建立永磁同步电机系统模型, 取得了较好的预测效果. 与整数 阶滤波器BiGRU-attention神经网络相比, 电压预测值的均方误差降低了30.87%, 平均绝对误差降低了26.97%; 电流 预测值的均方误差降低了34.42%, 平均绝对误差降低了14.88%.

英文摘要

      In this paper, modeling and identification method of Wiener nonlinear system based on fractional order filter and bi-directional gated recurrent unit (BiGRU) neural network is proposed, and applied to current and voltage prediction of permanent magnet synchronous motors. Firstly, the unknown coefficient of the fractional order filter is computed by the Grunwald-Letnikov method, and then the filter is utilized to filter the input data to remove the high frequency noise ¨ and enhance the robustness of the data. Secondly, in order to improve the ability to capture the deeper features of the sequences, the BiGRU neural network is used to simultaneously acquire the past and future information of the sequence data, and the parameters of the BiGRU network are updated by an adaptive momentum estimation technique. Simulation results show that the proposed Wiener system could effectively model the permanent magnet synchronous motor system and achieve better prediction results. Compared with the integer order filter BiGRU neural network, the mean square error of the voltage prediction value is reduced by 30.87% and the mean absolute error is reduced by 26.97%, the mean square error of the current prediction value is reduced by 34.42% and the mean absolute error is reduced by 14.88%.