Prediction of NOx concentration based on interpretable convolutional gated recurrent unit with clustering-extracting features

An interpretative convolutional gated recurrent unit with clustering-extracting features was proposed for the prediction of NOx concentration in SCR system. Firstly, a novel adaptive feature reconstruction approach utilizing hierarchical clustering and principal component extraction is presented for generating the input feature set of the model. This method involves adaptive feature clustering employing the “Ward” algorithm and feature set reconstruction based on cluster-oriented principal component analysis (PCA). Subsequently, a convolutional gated recurrent unit (ConvGRU) dynamic prediction model composed of one input layer and two hidden layers is constructed to predict the NOx concentration. Finally, the feature contribution mechanism is analyzed in model by Shapley additive explanations (SHAP). Application results on a 350 MW boiler with the front and rear wall opposed-firing demonstrate that the reconstructed feature set can reduce redundancy by 77.8 % while preserving clear physical significance of each feature. The R2 of the proposed modeling scheme on the test set is 0.9338, with MAE, RMSE, and MAPE values of 6.0462 mg/m3, 8.1158 mg/m3, and 3.274 %, respectively. Compared to the benchmark ConvGRU model, these metrics are reduced by 5.7 %, 9.1 %, and 8.03 %, respectively, while the training efficiency is enhanced by 49.03 %. SHAP analysis results indicate that oxygen parameter and global operating parameter of the unit significantly influence the outcomes of the prediction model. The modeling scheme proposed in this study satisfies the accuracy and interpretability requirements for NOx concentration prediction and offers a reference for predicting NOx concentrations.