Samples(25500)were collected from a selective catalytic reduction(SCR)denitrification system in a fluid catalytic cracking unit and preprocessed using the quartile method and the K-nearest neighbors interpolation meth...Samples(25500)were collected from a selective catalytic reduction(SCR)denitrification system in a fluid catalytic cracking unit and preprocessed using the quartile method and the K-nearest neighbors interpolation method to remove outliers.Using the Pearson correlation coefficient and LightGBM feature score method,13 key operational variables were identified and used to establish a model to predict outlet nitrogen oxide(NO_(x))concentration in an SCR system with backpropagation neural network,long short-term memory(LSTM)and LSTM-attention fully connected(FC)model,respectively.The LSTM-attention FC model showed better accuracy and generalization capability compared with other models.Its mean square error,mean absolute error,and coefficient of determination on the training and test datasets were 11.32 and 12.51,3.65%and 3.97%,and 0.96 and 0.94,respectively.Furthermore,a combination of the LSTM-attention FC model with a genetic algorithm used to optimize four feature variables including ammonia pressure compensation,inlet pressure,gas inlet upper temperature,and outlet ammonia concentration.The outlet NO_(x)concentration could be controlled below 80±3 mg/m^(3),and the ammonia slip concentration could be controlled below 0.1 mg/m^(3),demonstrating that the optimization model can provide effective guidance for reducing NO_(x)emissions and ammonia slip of SCR systems.展开更多
传统的石油化工过程建模中仅使用静态数据,而未能充分考虑连续生产过程中时序信息对建模指标的影响。本文提出了一种静态与时序数据组合网络(CNSS)模型,使用前馈神经网络提取静态数据的信息,使用Bi-LSTM(Bidirectional-Long Short Term ...传统的石油化工过程建模中仅使用静态数据,而未能充分考虑连续生产过程中时序信息对建模指标的影响。本文提出了一种静态与时序数据组合网络(CNSS)模型,使用前馈神经网络提取静态数据的信息,使用Bi-LSTM(Bidirectional-Long Short Term Memory)和自注意力机制提取操作变量时序数据中的信息,其中Bi-LSTM提取操作变量在时序逻辑上的信息,自注意力机制提取操作变量之间的交叉信息,通过静态和时序数据信息的组合以获得更好的模型预测性能;并使用CNSS模型分别对S Zorb装置精制汽油辛烷值(RON)、催化裂化烟气脱硝系统氮氧化物(NO_(x))的出口质量浓度进行预测,结果表明:CNSS模型的预测精度明显高于仅使用静态数据的机器学习模型,其对精制汽油RON预测的平均绝对误差和平均绝对百分比误差分别为0.1091、0.12%,对NO_(x)出口质量浓度预测的平均绝对误差和平均绝对百分比误差分别为2.4430 mg/m3、5.60%。对于因工艺参数波动较大而需要考虑时序信息的石油化工过程,CNSS模型可以为其建立机器学习模型提供重要参考。展开更多
基金This work was supported by the SINOPEC:Development of Remote Diagnosis Technology for FCC Flue Gas Desulfurization and Denitrification(320076).
文摘Samples(25500)were collected from a selective catalytic reduction(SCR)denitrification system in a fluid catalytic cracking unit and preprocessed using the quartile method and the K-nearest neighbors interpolation method to remove outliers.Using the Pearson correlation coefficient and LightGBM feature score method,13 key operational variables were identified and used to establish a model to predict outlet nitrogen oxide(NO_(x))concentration in an SCR system with backpropagation neural network,long short-term memory(LSTM)and LSTM-attention fully connected(FC)model,respectively.The LSTM-attention FC model showed better accuracy and generalization capability compared with other models.Its mean square error,mean absolute error,and coefficient of determination on the training and test datasets were 11.32 and 12.51,3.65%and 3.97%,and 0.96 and 0.94,respectively.Furthermore,a combination of the LSTM-attention FC model with a genetic algorithm used to optimize four feature variables including ammonia pressure compensation,inlet pressure,gas inlet upper temperature,and outlet ammonia concentration.The outlet NO_(x)concentration could be controlled below 80±3 mg/m^(3),and the ammonia slip concentration could be controlled below 0.1 mg/m^(3),demonstrating that the optimization model can provide effective guidance for reducing NO_(x)emissions and ammonia slip of SCR systems.
文摘传统的石油化工过程建模中仅使用静态数据,而未能充分考虑连续生产过程中时序信息对建模指标的影响。本文提出了一种静态与时序数据组合网络(CNSS)模型,使用前馈神经网络提取静态数据的信息,使用Bi-LSTM(Bidirectional-Long Short Term Memory)和自注意力机制提取操作变量时序数据中的信息,其中Bi-LSTM提取操作变量在时序逻辑上的信息,自注意力机制提取操作变量之间的交叉信息,通过静态和时序数据信息的组合以获得更好的模型预测性能;并使用CNSS模型分别对S Zorb装置精制汽油辛烷值(RON)、催化裂化烟气脱硝系统氮氧化物(NO_(x))的出口质量浓度进行预测,结果表明:CNSS模型的预测精度明显高于仅使用静态数据的机器学习模型,其对精制汽油RON预测的平均绝对误差和平均绝对百分比误差分别为0.1091、0.12%,对NO_(x)出口质量浓度预测的平均绝对误差和平均绝对百分比误差分别为2.4430 mg/m3、5.60%。对于因工艺参数波动较大而需要考虑时序信息的石油化工过程,CNSS模型可以为其建立机器学习模型提供重要参考。
