NOx concentration prediction in thermal power units based on SHAP interpretability analysis: A deep learning model VSAttLSTM with multimodal data

Abstract

Accurate prediction of NOx concentration is of significant importance for pollutant emissions and safe operation in thermal power plants. However, a single data-driven model cannot describe the global properties of the research object, which hinders generalization performance and introduces uncertainty. To address this issue, this paper proposes a deep learning prediction model, VSAttLSTM, which integrates a signal decomposition algorithm and a hyperparameter adaptive optimization algorithm. The model combines Long Short-Term Memory (LSTM) networks with a self-attention mechanism, and introduces the SSA (Sparrow Search Algorithm) for hyperparameter optimization to enhance the model's predictive performance and stability. This study also employs the SHAP (Shapley Additive Explanations) method to conduct interpretability analysis on the benchmark deep learning model, revealing the extent of influence of various features on the prediction results of NOx concentration. The variational mode decomposition (VMD) algorithm is applied to decompose the NOx pollutant data for signal decomposition, enabling multi-task modeling. The results of comparative experiments and ablation experiments demonstrate that the VSAttLSTM model significantly improves prediction accuracy, enabling the relative prediction error of the original data to be controlled within 5 %, and outperforms the comparative models on multiple performance metrics. This provides the industry with a more transparent and practical method for predicting NOx pollutant emissions from thermal power units.