引用本文:方小燕,姚立忠,罗海军,张玉泽,易军.基于自适应双层无迹卡尔曼滤波神经网络的铝电解电流效率预测模型[J].控制理论与应用,2025,42(3):579~589.[点击复制]
FANG Xiao-yan,YAO Li-zhong,LUO Hai-jun,ZHANG Yu-ze,YI Jun.Prediction model for the current efficiency of aluminum electrolysis based on the adaptive double layer unscented Kalman filter neural network[J].Control Theory and Technology,2025,42(3):579~589.[点击复制]
基于自适应双层无迹卡尔曼滤波神经网络的铝电解电流效率预测模型
Prediction model for the current efficiency of aluminum electrolysis based on the adaptive double layer unscented Kalman filter neural network
摘要点击 62  全文点击 3  投稿时间：2023-02-06  修订日期：2025-03-15
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2024.30047
  2025,42(3):579-589
中文关键词  铝电解  自适应建模  双层无迹卡尔曼滤波  人工神经网络  电流效率
英文关键词  aluminum electrolysis  adaptive modeling  double layer unscented Kalman filtering  neural network  current efficiency
基金项目  国家自然科学基金项目(62373069, 51805059), 重庆市教委科学技术研究项目(KJQN202200531, KJZD–M202101501), 重庆师范大学基金项目 (22XLB014)资助.
作者单位E-mail
方小燕 重庆科技学院 fangxiaoyan1009@163.com 
姚立忠* 重庆师范大学 lizhong_yao@cqnu.edu.cn 
罗海军 重庆国家应用数学中心  
张玉泽 重庆科技学院  
易军 重庆科技学院  
中文摘要
      针对铝电解过程强干扰和强时变导致模型精确度和稳定性不佳的难题,本文提出一种基于自适应双层无迹卡尔曼滤波神经网络的建模方法.该方法首先构建一种双层无迹卡尔曼滤波神经网络模型,以提高模型对扰动系统的稳定性.具体为:使用双层无迹卡尔曼滤波在线更新神经网络的权值和阈值;然后,在双层无迹卡尔曼滤波神经网络的状态变量均方误差中引入约束调节参数;同时,采用梯度下降法自适应调整比例调节参数,将其均方误差约束至较小的范围内,以此来削弱滤波递归计算过程中误差累积对模型的影响;最后,通过铝电解电流效率预测,验证了本文所提方法具有较高的精确度和稳定性.
英文摘要
      This paper presents a novel modeling method based on an adaptive double layer unscented Kalman filter neural network (ADLUKFNN), which tackles the challenges of poor model accuracy and stability resulting from strong interference and time-varying disturbances in the aluminum electrolysis process. Firstly, this method constructs a double layer unscented Kalman filter neural network (DLUKFNN) model to enhance the stability of the model towards the disturbance system. Specifically, the weights and thresholds of the neural network are updated online using the double layer unscented Kalman filter. Then, a constraint adjustment parameter is introduced into the mean square error of the state variables in DLUKFNN. Meanwhile, by employing the gradient descent method to adaptively adjust the constraint adjustment parameter, the mean square error is constrained within a smaller range, thereby weakening the impact of error accumulation during the filtering recursive calculation on the model. Finally, the accuracy and stability of the proposed method are verified through aluminum electrolysis current efficiency prediction.