基于小波去噪和循环神经网络-k重-整合移动平均自回归模型的转炉煤气柜位预测  

Prediction of converter gas tank levels based on wavelet threshold denoising method and RNN-k-ARIMA mixed model

作　　者：钱金花[1] 郑文娟 吴文彬 徐晨[1] QIAN Jinhua;ZHENG Wenjuan;WU Wenbin;XU Chen(College of Sciences,Northeastern University,Shenyang 110819,China;Intelligent Manufacture Center,Fushun New Steel Co.,Ltd.,Fushun 113001,China)

机构地区：[1]东北大学理学院,辽宁沈阳110819 [2]抚顺新钢铁有限责任公司智造中心,辽宁抚顺113001

出　　处：《冶金自动化》2023年第3期24-34,共11页Metallurgical Industry Automation

基　　金：国家自然科学基金青年科学基金项目(11801065)。

摘　　要：在钢铁企业的生产过程中,实时预测钢铁企业的转炉煤气柜位对转炉煤气系统的优化调度至关重要。本文提出基于小波去噪和循环神经网络-k重-整合移动平均自回归模型(recurrent neural network-k-autoregres-sive integrated moving average model,RNN-k-ARIMA)混合预测模型,并用该模型预测转炉煤气柜位。其主要研究思路为:首先,利用小波阈值去噪方法去除柜位数据中的干扰噪声;其次,运用RNN模型训练去噪后的柜位数据,计算RNN模型预测结果与实际值的残差;再次,使用k-ARIMA模型对残差进行修正;最后,将修正后的残差与RNN模型预测的初值求和得到最终的预测结果。通过测试两个数据集,得到均方根误差ERMS(root mean square error,RMSE)分别为0.206142和0.146249,平均绝对百分比误差EMAP(mean absolute percentage error,MAPE)分别为0.941101和0.720312,方向精度A_(D)(directional accuracy,DA)均为0.833333。综合对比发现,相比于单独应用RNN模型与ARIMA模型,使用RNN-k-ARIMA混合模型预测转炉煤气柜位的精度更高、性能更好,弥补了传统时间序列模型和单一网络预测模型的不足。该混合预测模型可以为钢铁企业转炉系统的管理与调度提供科学的理论依据。In the production process,real time prediction of Lindz-Donawitz converter gas tank levels plays an important role in optimizing gas scheduling system for iron and steel enterprises.In order to improve the prediction accuracy of converter gas tank levels,a mixed prediction model named RNN-k-ARIMA was proposed to predict the Lindz-Donawitz converter gas tank levels in this paper.The re-search content was arranged as following.First of all,removing the interference noise in the tank level data by the aid of the wavelet threshold denoising method.Secondly,training the denoised tank level data by use of the RNN model and calculating the residual derived by comparing the predicted results from the RNN model to the actual value.Continuously,correcting the residual of the tank level data by the k-fold ARIMA model until the new residual sequence fails to pass the white noise test.Finally,summing the corrected residual with the initial value predicted by RNN model to obtain the final pre-diction results.By testing the two datasets,the ERMS are 0.206142 and 0.146249,the EMAP are 0.941101 and 0.720312,respectively,and the AD is 0.833333.The comprehensive comparison shows that the RNN-k-ARIMA mixed model appears higher accuracy and better performance in predicting the converter gas tank levels,which can make up for the deficiency of the traditional time series model and single network prediction model.The mixed prediction model can provide a scientific theoretical basis for the iron and steel enterprises to manage and schedule the Lindz-Donawitz gas systems.

关 键 词：小波去噪 RNN模型 k-ARIMA模型 混合预测模型 转炉煤气 

分 类 号：TF341[冶金工程—冶金机械及自动化] TP183[自动化与计算机技术—控制理论与控制工程]