基于CNN-LSTM的燃气轮机NO_(x)排放预测研究

Research on NO_(x) Emission Prediction of Gas Turbine based on CNN-LSTM

为了建立精准的NO_(x)预测模型,解决燃气轮机电站存在NO_(x)超标排放的问题,提出一种基于卷积神经网络(Convolutional Neural Network, CNN)和长短期记忆神经网络(Long Short-term Memory, LSTM)组合模型的NO_(x)排放预测方法。将NO_(x)排放历史数据和燃气轮机燃烧的状态参数通过滑动窗口法构建成特征图格式输入到CNN中,利用其卷积层和池化层提取表征NO_(x)动态变化的特征向量,并转化为时间序列格式输入到LSTM中进一步挖掘内部规律,从而实现NO_(x)的排放预测。以某三菱燃气轮机的历史运行数据进行试验。结果表明:CNN-LSTM的相对均方误差eRMSE为1.811 mg/m^(3),并通过与PCA-BP,PCA-RNN和PCA-LSTM模型进行比较,验证了方法的可行性。

In view of the current problem of excessive NO_(x) emissions in gas turbine power plants, the accurate prediction model of NO_(x) emission is the basis of reducing NO_(x) emission.Therefore, a new NO_(x) emission prediction method based on the combined model of CNN and LSTM is proposed.The historical NO_(x) emission data and the state parameters of gas turbine combustion are constructed into a feature map format by sliding window algorithm, which are input into CNN,its convolutional layer and pooling layer are used to extract feature vectors that characterize the dynamic changes of NO_(x),which are transformed into a time series format and then input into LSTM to further explore the internal rules, and realize the NO_(x) emission prediction.In the end, the historical operation data of a Mitsubishi gas turbine are used for the test.The result shows that the relative mean square error eRMSE of CNN-LSTM is 1.811 mg/m^(3),and the feasibility of this method is verified by comparing with PCA-BP,PCA-RNN and PCA-LSTM models.