| Abstrakt: |
Objective] Because of the limited arrangement of NOx emission measurement points in CFB units, the reductant injection amount is often inaccurate. Moreover, given that the pollutant generation characteristics of the units are different under different loads, higher requirements are placed on the accurate measurement of NOx emissions. To analyze the NOx emission prediction of circulating fluidized bed units, the Informer neural network model is used to model the NOx emission of a 350 MW supercritical circulating fluidized bed unit. [Methods] First, the overview of the circulating fluidized bed unit denitration system and the Informer neural network model theory were introduced. On this basis, the data of a circulating fluidized bed unit running continuously for 50 h were obtained as sampling data, 20 parameters related to NOx emissions were determined as input characteristic parameters of the prediction model, and the relevant parameter data were standardized. Second, the simulation experiment platform and model evaluation indicators were determined, the simulation experiment steps were clarified, and the simulation experiment flowchart was drawn. On this basis, simulation experiments on 6 different NOx emission long-sequence time series predictions with prediction lengths of 12, 24, 36, 48, 72, and 96 were conducted, and regression and error analyses were performed on the simulation results of the 6 long-sequence predictions. Finally, according to the same experimental plan and the same operating data, the Transformer, RNN, and LSTM models were used to predict NOx emissions, and the prediction results of the Informer neural network model were compared with the prediction results of the three models in terms of evaluation indicators and time consumption. The comparison results of the evaluation indicators of the four models were presented in a table, and the comparison results of time consumption were presented in a bar chart. [Results] Results showed that the Informer neural network model has good feature extraction and long-sequence input abilities through the attention and distillation mechanisms. The NOx emission prediction effect of this model is significantly better than those of the Transformer, RNN, and LSTM models in terms of prediction accuracy and time consumption. When the prediction length is 24, 36, 48, 72, and 96, the evaluation index value of the Informer neural network model is the smallest, and its prediction accuracy is better than those of the three comparison models. When the prediction length is 12, the sparse attention mechanism of the Informer neural network model cannot effectively extract the periodic characteristics of the input data. The prediction accuracy of the Informer neural network model is slightly worse than that of the LSTM model but is significantly better than that of the Transformer and RNN models. The average time consumption of the Informer neural network model is lower than those of the three comparison models, and the average time consumption is reduced by 60% compared with the RNN model, 84% compared with the LSTM model, and 94% compared with the Transformer model. [Conclusions] The Informer neural network model can provide effective technical support for the prediction of NOx emissions of circulating fluidized bed units. This research serves the innovation and practical education of students majoring in energy and power at the China University of Mining and Technology, helps train students' scientific research thinking, and provides a certain reference for the development of practical teaching of energy and power majors. [ABSTRACT FROM AUTHOR] |
