针对SCR系统出口NOx浓度测量准确性较差、CEMS测量存在较大延迟且吹扫过程中无法测量NOx浓度的问题，依托某660 MW机组的实际运行数据，建立基于长短期记忆神经网络(LSTM)的烟囱入口NOx浓度软测量模型。为了验证其对烟囱入口NOx浓度的预测性能，将所建LSTM模型与RNN模型、BPNN模型和KPLS 3种模型进行对比分析。研究表明：LSTM模型的预测均方根误差为1.34 mg/m3，平均相对误差为3.51%，模型预测精度优于其他3种模型；LSTM模型的泛化能力较强，数据动态跟踪效果好，具有较高的预测稳定性。

Aiming at the problems of poor accuracy of NO〖HT5〗x concentration measurement at SCR system outlet,large delay in CEMS measurement and inability to measure NO〖HT5〗x concentration during purging process,a soft measurement model of NO〖HT5〗x concentration at chimney inlet based on LSTM neural network was established based on the actual operation data of a 660 MW unit.In order to verify its predictive performance for NO〖HT5〗x concentration at chimney entrance,the LSTM model was compared with RNN model,BPNN model and KPLS model.The results show that the root mean square error of LSTM model is 1.34 mg/m3,and the average relative error is 3.51%.The prediction accuracy of LSTM model is better than the other three models.The LSTM model has strong generalization ability,good data dynamic tracking effect and high prediction stability.

TM621

上海市科委项目（14110502400）