基于改进LSTM的多时隙工业时序数据预测方法研究  

Research on Multi-Timeslot Industrial Timing Series Data Prediction Method Based on Improved LSTM

作　　者：周红福 孙凯文 ZHOU Hongfu;SUN Kaiwen(Shanghai Institute of Process Automation&Instrumentation Co.,Ltd.,Shanghai 200233,China)

机构地区：[1]上海工业自动化仪表研究院有限公司,上海200233

出　　处：《自动化仪表》2025年第4期86-91,共6页Process Automation Instrumentation

摘　　要：在工业生产中,常常存在对仪器仪表数据进行趋势预测的需求。对长短期记忆(LSTM)神经网络作出改进,提出一种多时隙工业时序数据预测方法。首先,对输入端作出改进,使得模型能够预测多个采样周期后的数值。其次,对模型单元作出优化,提高了模型对数值的拟合能力。再次,设计了一种新颖的数据清洗算法,提升了数据获取的稳定性。最后,使用污水厂的多组真实流量数据,对该方法进行验证。验证结果表明,该方法克服了原始LSTM预测方法的缺陷,创新地完成了多时隙数据预测的任务,实现了对24 h后数据的精准预测;相比对照组方法,该方法在数据曲线图像跟随趋势与数学统计指标方面均有提升。该方法能够实际助力污水厂的资源计划和调配。In industrial production,there is often a need for trend prediction of instrumentation data.By improving long shortterm memory(LSTM)neural network,a multi-timeslot industrial timing series data prediction method is proposed.Firstly,the inputs are improved,so that the model can predict the values after multiple sampling periods.Secondly,the model units are optimized to improve the model's ability to fit the values.Thirdly,a novel data cleaning algorithm is designed to improve the stability of data acquisition.Finally,the method is validated using multiple sets of real flow data from wastewater plants.The validation results show that the method overcomes the defects of the original LSTM prediction method,innovatively accomplishes the task of multi-timeslot data prediction,and realizes the accurate prediction of the data after 24 h;compared with the methods in the control group,the method has improvement in aspects of the trend of following the data curve image and mathematical statistical indexes.The method can practically help the resource planning and deployment of wastewater plants.

关 键 词：长短期记忆神经网络 深度学习 趋势预测 工业时序数据 滑动窗口 数据清洗 

分 类 号：TH17[机械工程—机械制造及自动化]