计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (20): 277-285.DOI: 10.3778/j.issn.1002-8331.2103-0292

• 工程与应用 • 上一篇    下一篇

互信息深度稀疏自编码融合DLSTM预测网络

李江坤,黄海燕   

  1. 华东理工大学 化工过程先进控制和优化技术教育部重点实验室,上海 200030
  • 出版日期:2022-10-15 发布日期:2022-10-15

Mutual Information Deep Sparse Auto-Encoding Hybrid DLSTM Prediction Network

LI Jiangkun, HUANG Haiyan   

  1. Key Laboratory of Advanced Control and Optimization for Chemical Processes Ministry of Education, East China University of Science and Technology, Shanghai 200030, China
  • Online:2022-10-15 Published:2022-10-15

摘要: 针对工业系统变量之间存在动态和相互关联特性导致关键变量预测精度降低问题,提出一种互信息深度堆叠稀疏自编码数据特征网络(mutual information-deep stack sparse auto-encoder,MI-DSSAE)结合深度长短期记忆(deep LSTM,DLSTM)预测模型。MI-DSSAE模型对稀疏编码器改进,采用堆叠稀疏编码器结构,引入互信息作为重构损失权重,对多个稀疏编码器隐层迁移并微调。预测部分采用深度DLSTM网络结构,用双层Bi-LSTM结构对序列数据的动态变化特性双向捕捉,将输出数据输入到普通LSTM进行记忆处理,进行全连接层加权预测关键质量变量。采用流程化工业案例脱丁烷塔的C4含量对提出的模型验证,同时对比RNN、LSTM、GRU模型以及MI-DSSAE-RNN、MI-DSSAE-LSTM、MI-DSSAE-GRU等模型,通过RMSE、R2和MAE多项回归误差指标对比分析,验证MI-DSSAE-DLSTM模型的有效性。

关键词: 深度稀疏自编码器, 互信息, 长短期记忆(LSTM), 脱丁烷塔, 软测量

Abstract: In the view of dynamic and interrelated characteristics in industrial system variables, which leads to the problem of reduced prediction accuracy of key variables, a mutual information-deep stack sparse auto-encoder(MI-DSSAE) combined with deep long short-term memory(DLSTM) prediction model is proposed. MI-DSSAE improves the sparse encoder and superimposes the hidden layer into a stacked sparse encoder structure, introducing mutual information as the reconstruction loss weight, and then migrates and fine-tunes multiple sparse encoders. The prediction part adopts the deep two-way DLSTM network structure. Firstly, the two-layer Bi-LSTM structure is used to capture the dynamic characteristics of the sequence data in two directions, and the output data is input into the ordinary LSTM for memory processing. Finally, the fully connected layer is weighted to predict the key quality variables. The C4 content change of the debutanizer column in the process industry case is used to verify the proposed modeling, while RNN, LSTM, GRU, MI-DSSAE-RNN, MI-DSSAE-LSTM, and MI-DSSAE-GRU methods are compared with proposed method. The comprehensive comparative analysis of multiple regression error indicators of RMSE, R2, and MAPE shows the effectiveness of the MI-DSSAE-DLSTM modeling method.

Key words: deep sparse auto-encoder, mutual information, long short-term memory(LSTM), debutanizer column, soft measurement