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
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.
