基于MTCN-Informer的铁矿球团工艺预测模型  

Prediction Model of Iron Ore Pellet Process Based on MTCN-Informer

作　　者：廖雪超[1,2] 朱晨辉 赵昊裔 向桂宏 刘宗宇 LIAO Xue-chao;ZHU Chen-hui;ZHAO Hao-yi;XIANG Gui-hong;LIU Zong-yu(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System,Wuhan 430065,China;Wisdri City Construction Engineering&Research Incorporation Ltd,Wuhan 430062,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]智能信息处理与实时工业系统湖北省重点实验室,湖北武汉430065 [3]中冶南方城市建设工程技术有限公司,湖北武汉430062

出　　处：《计算机技术与发展》2024年第9期188-194,共7页Computer Technology and Development

基　　金：国家自然科学基金项目(62273264)。

摘　　要：成品球团流量的预测是生产过程的关键,它决定着整个生产的效率和产量。铁矿球团链箅机—回转窑是生产铁矿石制备高品质铁合金的重要工艺过程之一,具有大时滞、参数庞杂、耦合关系复杂等特点,且成品球团流量波动剧烈,使球团流量难以预测。为此,该文使用移动平均滤波器来平滑波动的数据,互信息法对庞杂的参数做特征选择,再利用基于自注意力机制的Informer球团流量预测模型,其降低传统自注意力机制的时间复杂度,提高了模型训练效率。同时,针对Informer模型的概率稀疏自注意力机制难以把握长时间序列波动的问题,通过TCN时间卷积网络来提取长时间序列的扩展信息依赖,同时结合Informer编码解码网络来处理上下文的信息,从而完成球团流量的精确预测。通过对工厂实际数据进行实验分析可知,与循环神经网络这类传统的深度学习模型相比,所提集成模型在预测精度、稳定性方面均为最优。The prediction of finished pellet flow is the key to the production process,which determines the efficiency and output of the whole production.Iron ore pellet chain grate machine—rotary kiln is one of the important processes for producing iron ore to prepare high-quality ferroalloy.It has the characteristics of large time lag,complex parameters,complex coupling relationship,etc.,and the flow rate of finished pellets fluctuates violently,making the flow rate of pellets difficult to predict.For this reason,we use the moving average filter to smooth the fluctuating data,and the mutual information method performs feature selection on complex parameters,and then uses the Informer pellet flow prediction model based on the self-attention mechanism,which reduces the time of the traditional self-attention mechanism complexity and improves the efficiency of model training.At the same time,in view of the problem that the probabilistic sparse self-attention mechanism of the Informer model is difficult to grasp the long-term sequence fluctuations,the extended information dependence of the long-term sequence is extracted through the TCN time convolution network,and the context information is processed by combining the Informer encoding and decoding network,thereby completing accurate prediction of pellet flow.Through the experimental analysis of the actual factory data,it can be seen that compared with traditional deep learning models such as recurrent neural networks,the proposed integrated model is the best in terms of prediction accuracy and stability.

关 键 词：球团流量预测 特征选择 时间卷积网络 编码解码网络 自注意力机制 

分 类 号：TP390[自动化与计算机技术—计算机应用技术]