Dynamic prediction of boiler NOx emission with graph convolutional gated recurrent unit model optimized by genetic algorithm

A novel approach for the dynamic prediction of boiler NOx is proposed that integrates the graph convolutional network (GCN) with the gated recurrent unit (GRU) model, referred to as GC-GRU model. The operational status of boiler parameters is treated as graph-structured data, and the GCN is used to extract the spatial characteristics of boiler parameters which will then be taken as the input of the GRU that considers the dynamic characteristics of boiler operation. The GC-GRU model can simultaneously take into consideration the spatio-temporal characteristics of boiler NOx to realize accurate prediction of boiler NOx. This study presents a complete procedure for boiler NOx prediction ranging from the selection of key feature variables for model input to the optimization of model hyperparameters. Specifically, the MIC analysis is used to extract the key boiler parameters that are closely associated with NOx emission. The Genetic Algorithm (GA) is proposed to optimize the model hyperparameters (GA-GC-GRU model). The results show that the GA-GC-GRU model exhibits superior performance because of its ability to capture the spatio-temporal characteristics of boiler NOx. Specifically, the test MAPE of the model is 2.41%, reduced by more than 25% compared to those of SVM, DNN, and LSTM models.