双向门控循环神经网络的SO_(2)排放浓度预测模型  

Prediction model of SO_(2) emission concentration based on bidirectional gated recurrent unit neural network

作　　者：蒋星明 曹顺安[1] 王民军 陈东[1] 董毕承 JIANG Xing-ming;CAO Shun-an;WANG Min-jun;CHEN Dong;DONG Bi-cheng(School of Power and Mechanical Engineering,Wuhan University,Wuhan 430072,China;Zhejiang Energy and China Coal Energy Zhoushan Coal Power Co.,Ltd.,Zhoushan 316000,China)

机构地区：[1]武汉大学动力与机械学院,湖北武汉430072 [2]浙江浙能中煤舟山煤电有限责任公司,浙江舟山316000

出　　处：《应用化工》2021年第12期3519-3523,共5页Applied Chemical Industry

基　　金：国家自然科学基金资助项目(21906123)。

摘　　要：火力发电站脱硫系统数据具有大惯性和延时性等特点,且影响SO_(2)排放浓度的因素众多。为此,建立了基于双向门控循环神经网络(biGRU)的SO_(2)排放浓度预测模型。以分析得到的主成分为输入变量,SO_(2)排放浓度为输出变量,通过训练对脱硫系统SO_(2)排放浓度数据进行预测,并进行比较。结果表明,与传统的RNN以及LSTM模型相比,biGRU模型能够获得较高的预测精度,其对称平均绝对百分比误差相较于RNN和LSTM分别下降了4.235%,0.718%,其均方根误差分别下降了1.942,0.443 mg/Nm^(3)。该模型预测误差较低,泛化能力较好,具有较高的实际应用价值,有利于实现排放控制和节能减排。Aiming at the characteristics of large inertia,time delay and many influencing factors of SO_(2)emission concentration in the desulfurization system data,a prediction model of SO_(2)emission concentration based on bidirectional gated recurrent unit neural network(biGRU)is proposed.Taking the main components obtained from the analysis as the input variables and the SO_(2)emission concentration as the output variable,the SO_(2)emission concentration is predicted through the data training.The comparison reveals that the biGRU model can obtain higher prediction accuracy.Compared with RNN and LSTM,the symmetric mean absolute percentage error decreased respectively by 4.235%and 0.718%,and the root mean square error decreased respectively by 1.942 mg/Nm^(3) and 0.443 mg/Nm^(3),which indicates that the biGRU model has lower prediction error,better generalization ability and higher practical application value,and is conducive to protect the environmental protection and reduce emission.

关 键 词：火力发电站 脱硫系统 SO_(2)排放浓度 biGRU 预测模型 

分 类 号：TQ015.9[化学工程]