Attention based spatial-temporal graph convolutional networks for boiler NOx prediction

Due to the introduction of large-scale renewable energy to the electric grid, the coal-fired units are running more under load cycling conditions and this has dramatically increased the difficulty of boiler in the control of NOx emissions. The real-time prediction of NOx emissions is of great significance to guide the efficient and clean operation of thermal power units under load cycling conditions. The CFD methods require huge computational cost to solve multiple coupled conservation equations, and hence, are infeasible to establish the complex nonlinear mapping relationship between boiler operating parameters and NOx emissions in a real-time manner. With the rapid development of big data analysis and artificial intelligence, data-driven models provide a new approach to realize the real-time prediction and control of boiler NOx emissions. Under load cycling conditions the boiler operating data has strong correlation in time sequence and complex spatial correlation characteristics. However, the traditional neural network methods cannot effectively capture these correlation characteristics, thus, greatly limiting their ability in the accurate prediction of boiler NOx emission. To resolve this problem, a boiler NOx prediction model based on the spatiotemporal attention graph convolutional network (AST-GCN) is proposed in this study. Not only can this model capture the spatial correlations among the boiler operating parameters, but also can it capture the dynamic relationship between the historical boiler operating data and NOx emissions. In addition, the attention mechanism embedded in the model can further improve the ability to adaptively extract the spatio-temporal correlation features from the boiler operating data, and increase the interpretability of the model. The prediction results based on the operating data of a 600 MW boiler demonstrate that the AST-GCN model exhibits a greatly improved prediction accuracy and generalization performance since it can effectively capture the spatial and time sequence correlations among the boiler operating parameters in comparison to the traditional neural network models.