【摘要】 催化裂化装置工艺复杂,调整工艺参数极易发生连锁反应,采用传统的集总模型对污染排放进行预测的难度较大。针对炼化企业海量生产数据和污染排放数据多参数多变量相互耦合的特点,利用主成分分析(PCA,Principal Component Analysis)对NO_x排放要素进行特征选择,确定原料中氮含量、反应温度、剂油比、停留时间等为关键生产要素;基于长短期记忆(LSTM,Long Short-Term Memory)网络建立NO_x排放预测模型,对某350万t重油催化裂化装置NO_x排放进行预测,与卷积神经网络(CNN,Convolutional Neural Networks)、支持向量机(SVM,Support Vector Machine)以及BP神经网络(Back Propagation Neural Networks)进行了对比分析。结果表明,由于考虑了时间序列内部的数据特性,LSTM的平均绝对误差、均方根误差、皮尔逊相关系数和可决系数等指标均优于其他方法。更多还原

【Abstract】 Because of the complexity of FCC process,to adjust its process parameters is very easy to cause chain reaction.It is difficult to predict the pollution emission of FCC processby using the traditional lumped model.In order to solve the problem of mutual coupling of multi parameters in the mass production data and pollution emission data of refining and chemical enterprises,the main influence factors of NO_x emission are determined using principal component analysis(PCA),and they are nitrogen content in raw material,reaction temperature,ratio of catalyzer to crude oil and residence time of raw material.NO_x emission prediction model is established using long-short term memory(LSTM) network,and the NO_x emission of a 3.5 million ton RFCC unit is predicted using it.The prediction result of LSTM network is compared with those of CNN,SVM and BP neural networks.It is shown that,considering the internal data characteristics of time series,the mean absolute error,root mean square error,Pearson correlation coefficient and determinable coefficient of the prediction result of LSTM network are better than those of CNN,SVM and BP neural networks.更多还原