Performance and Statistical Assessment of Deep Learning Frameworks for Low-Temperature Selective Catalytic Reduction of NO x

Abstract

Nitrogen oxides (NO x ) are among the most harmful emissions produced by internal combustion engines, contributing to air pollution, acid rain, and climate-related impacts. Accurate modeling of the NO x reduction process is therefore essential for developing efficient emission control technologies and optimizing catalytic performance. This study explores the potential of advanced deep learning architectures to predict NO x conversion efficiency in an ethanol-assisted selective catalytic reduction (SCR) system. Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and a hybrid TCN-LSTM were designed and optimized using Bayesian optimization to obtain the most effective hyperparameter configuration. Experimental data collected from a two-cylinder diesel engine equipped with Ag-Sn-P-based catalysts were used for model training and validation. Model performance was evaluated through Mean Absolute Error (MAE) and Theil's U2 indicators, supported by bootstrap confidence interval analysis to ensure statistical robustness. Among the tested models, the TCN achieved the best predictive accuracy, with a mean absolute error of 0.118 and a Theil's U2 value of 0.051, followed closely by the hybrid TCN-LSTM model with an MAE of 0.132 and U2 of 0.053. The results confirm that convolutional and hybrid architectures provide reliable and generalizable frameworks for capturing the nonlinear dynamics of catalytic NO x reduction, offering valuable insights for next-generation emission control strategies.