针对燃煤机组选择性催化还原(SCR)系统出口氮氧化物（NOx）预测模型精度不高的问题，提出一种基于最大信息系数（MIC）和长短期记忆（LSTM）神经网络的预测模型方法。首先采用MIC估计各变量的延迟时间，对数据进行时延重构；然后采用重构后数据的MIC值作为评价各输入变量和输出变量间相关性大小的指标，并结合基于关联性的特征选择算法（CFS）进行输入变量筛选；最后基于时延重构和变量筛选后的数据，采用LSTM神经网络建立了SCR出口氮氧化物浓度动态预测模型。该模型被用于广东某320 MW燃煤机组实际运行数据分析。结果表明，经时延重构和变量筛选后所建立的LSTM预测模型具有较高精度，优于深度神经网络（DNN）模型和径向基函数（RBF）神经网络模型，平均绝对百分比误差达2.58%，均方根误差达2.02，可满足现场运用要求。

Aiming to improve the prediction accuracy of the concentration of nitrogen oxides (NOx) in the flue gas at the outlet of selective catalytic reduction (SCR) system for coal-fired power plants, a prediction model method based on the maximum information coefficient (MIC) and long-short term memory (LSTM) neural network was proposed. Firstly, MIC was used to estimate the delay time between various input parameters and the recorded NOx concentration, and the data were reconstructed according to the estimated delay time. Then the MIC value of the reconstructed data was used as an index to evaluate the correlation between input variables and output variables, and the correlation-based feature selection (CFS) algorithm was used to select the input variables. Finally, based on the data after time delay reconstruction and variable selection, the dynamic prediction model of NOx concentration at SCR outlet was established using LSTM neural network. The model was used to analyze the recorded operation data of a 320 MW coal-fired unit in Guangdong. The results show that the LSTM prediction model established after time delay reconstruction and variable selection has high accuracy, superior to deep neural networks (DNN) model and radial basis function (RBF) model, with the mean absolute percentage error of 2.58% and the root mean square error of 2.02, which can meet the requirements of field application. 

0 引言

1 基本原理

   1.1 最大信息系数

   1.2 长短期记忆神经网络

2 MIC-CFS-LSTM预测模型

   2.1 原始数据获取及数据预处理

   2.2 基于MIC的延迟时间分析

   2.3 基于MIC-CFS的特征变量选择

   2.4 LSTM模型参数确定

   2.5 模型建立整体流程

3 试验结果与分析

   3.1 延迟时间对预测精度的影响

   3.2 变量挑选对预测精度的影响

   3.3 不同预测模型对预测精度的影响

4 结论