针对电站锅炉在实际运行过程中存在燃烧优化调整不及时以及烟气脱硝成本较高的问题，提出了基于MI-GA-LSTM的炉膛出口NOx排放快速预测模型。根据燃煤锅炉实际运行数据，利用互信息（MI）进行特征相关性分析，将所得最优特征子集作为长短时记忆神经网络（LSTM）的输入，并利用遗传算法(GA)对模型关键参数进行寻优，得到炉膛出口NOx原始生成质量浓度预测的MI-GA-LSTM模型，并与LSTM、门控神经网络(GRU)、循环神经网络(RNN)在同一测试集上进行预测效果对比。结果表明：该模型在训练集和测试集上都能够对运行数据进行精准地预测，可以很好地完成多变量非线性拟合；该模型在测试集上的3项指标均优于其他模型，具有更高的预测精度和泛化能力；该模型可作为炉膛出口NOx排放质量浓度传感器的补充，提前准确感知炉膛出口NOx原始生成质量浓度的变化。

Considering the problems of untimely combustion adjustment and high cost of flue gas denitration in the actual operation of utility boilers, a fast prediction model of NOx emission at the outlet of furnace based on MI-GA-LSTM was proposed. Basing the realtime operation data of a coalfired boiler, mutual information (MI) was used to analyze the correlation, then the result of optimal variable quantum set was taken as the input of the long and shortterm memory (LSTM), and the genetic algorithm (GA) was applied to optimize the key parameters of the model to obtain the MI-GA-LSTM model for original generating mass concentration of NOx emission prediction at the outlet of furnace. The prediction effect of MI-GA-LSTM model was compared with that of LSTM, gated recurrent unit (GRU) and the recurrent neural network (RNN) prediction models in the same test data set. The results show that the MI-GA-LSTM model can accurately predict the NOx concentration on both the training sample data set and the test sample data set, and complete the multivariable nonlinear fitting well; the model is superior to other models with the three indices of the test set, and has higher prediction accuracy and generalization ability; the model can be used as a supplement to the NOx mass concentration sensor to precisely detect the change of the original generating mass concentration of NOx emission at the furnace outlet early.

TK32

贵州省科技支撑计划项目(\[2020\]2Y040);贵州西能电力建设有限公司科技创新项目（138021QT0320220012）