AG-STN: An Adaptive Graph Spatio-Temporal Network for NOx concentration prediction in coal-fired power plants

Accurate prediction of NO x concentrations in coal-fired power plants is crucial for environmental regulation and combustion optimization. However, prediction accuracy is often constrained by two major challenges: the non-stationarity of monitoring data and the complex dynamic coupling relationships among variables. To address these challenges, this paper proposes an Adaptive Graph Spatio-Temporal Network (AG-STN) model. The model first employs Complementary Ensemble Empirical Mode Decomposition (CEEMD) to decompose non-stationary raw sequences into more stable multi-scale components, thereby constructing information-rich graph node features. Subsequently, to characterize evolving inter-variable relationships under local operating conditions, the model employs the Sparrow Search Algorithm (SSA) to adaptively optimize the fusion weights between linear (PCC) and nonlinear (MI) correlations, thereby constructing window-specific fused graph adjacency matrices. Finally, the window-specific fused graphs and enhanced node features are integrated into a Graph Convolutional Network (GCN) and Long Short-Term Memory (LSTM) architecture to capture higher-order spatio-temporal dependencies. Experimental results based on real operational data from a coal-fired power unit show that the proposed model achieves the best overall forecasting performance among the compared methods, with an R 2 of 0.980. Additional experiments further verify the effectiveness of the proposed framework and its robustness under varying operating conditions.