基于STA-BiLSTM-LightGBM组合模型的多步预测软测量建模  

Multi-step predictive soft sensor modeling based on STA-BiLSTM-LightGBM combined model

作　　者：闫琳琦 王振雷[1] YAN Linqi;WANG Zhenei(Key Laboratory of Smart Manufacturing in Energy Chemical Process,Ministry of Education,East China University of Science and Technology,Shanghai 200237,China)

机构地区：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237

出　　处：《化工学报》2023年第8期3407-3418,共12页CIESC Journal

基　　金：国家重点研发计划项目(2021YFB3301303);国家自然科学基金面上项目(61973124,22178103);中央高校基本科研业务费专项;浦东新区科技发展基金项目(PKX2021-R03)。

摘　　要：在复杂工业生产过程中,为提高产品质量,建立关键变量多步预测模型非常必要,但传统软测量建模方法难以聚焦工业数据复杂特性,导致预测不准。本文提出一种基于时空注意力机制的双向长短时记忆网络与轻量级梯度提升机(spatial-temporal attention mechanism bi-directional long short-term memory network and light gradient boosting machine,STA-BiLSTM-LightGBM)的多步预测软测量模型。首先训练STA-BiLSTM,时空注意力机制从时间和空间维度为输入特征分配权重,BiLSTM捕捉数据时序特征;其次使用BiLSTM最后一个时间步的隐状态扩充原始输入数据后,训练LightGBM,利用弱学习器迭代训练得到最优模型;进而将STA-BiLSTM和LightGBM的预测输出按照误差倒数法变权求和得到预测结果。最后将该方法在工业数据集上仿真验证,结果表明组合模型预测效果优于BiLSTM和LightGBM,且随着预测步数增大,仍保持较高的预测精度。In complex industrial production processes,it is necessary to establish a multi-step forecasting model for key variables in order to improve product quality,but traditional soft sensor modeling methods are difficult to focus on the complex characteristics of industrial data,resulting in inaccurate forecasting.This paper proposes a multi-step predictive soft sensor model,which is the combination of the bi-directional long short-term memory network based on spatial-temporal attention mechanism and light gradient boosting machine(STA-BiLSTM-LightGBM).Firstly,the STA-BiLSTM is trained,meanwhile the spatial-temporal attention mechanism assigns weights to the input features according to the temporal and spatial dimensions,and the BiLSTM captures the temporal features of the data.Secondly,the implicit state of the last time step of BiLSTM is used to extend the original input data,and then LightGBM is trained.By training LightGBM with a weak learner,the optimal model can be obtained through iterative training.The predicted outputs of STA-BiLSTM and LightGBM can then be weighted to obtain the predicted results using the error reciprocal method.Finally,the simulation results demonstrate that the combined model is superior to both BiLSTM and LightGBM,and it maintains high prediction accuracy even as prediction steps increases.

关 键 词：软测量 预测 注意力机制 神经网络 组合模型 实验验证 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]