Intelligent ensemble architecture for capturing transient nitrogen oxides emission spikes in real-world driving conditions

Accurate prediction and monitoring of tailpipe nitrogen oxides (NOx) emissions is crucial for the design of effective automotive emission control strategies and for air quality in urban climates . This paper introduces a deep-learning soft sensing methodology designed to estimate tailpipe NOx emissions with high accuracy. The proposed method can capture both fast and slow transient behaviors of emissions, leading to accurate predictions of high spikes and smaller fluctuations. The central aspect of the proposed methodology is a temporal convolutional network (TCN) based classifier that differentiates between high and low NOx values. This differentiation allows the method to select the appropriate prediction model, specifically a TCN-based multi-head self-attention neural network . A hybrid methodology combining physics-based and data-driven approaches is employed for input feature selection, enhancing model robustness. The proposed method has been validated using an extensive dataset ( 2300km) of field data from the City of Edmonton, Canada transit buses. Comparative analysis with other prediction models, including deep neural networks (DNN), long-short-term memory networks (LSTM), TCN, and state-of-the-art TCNs with multi-head self-attention, demonstrate the advantage of the proposed approach. Notable improvements in capturing both high and low NOx transient dynamics compared to a DNN model are achieved. Specifically, the proposed model achieves a 78% improvement in mean absolute error (MAE) for high NOx values (from 21.77 to 4.67 [milligrams per second (mg/s)]) and a 59% improvement (from 0.72 to 0.29 [mg/s]) for low NOx values compared to the DNN models. For mean square error (MSE), the improvements are 89% (from 521.85 to 54.39 [mg/s] ) for high NOx values and 33% (from 1.47 to 0.98 [mg/s] ) for low NOx values compared to those in the DNN models. In terms of root mean square error (RMSE), the proposed method shows a 67% (from 22.85 to 7.37 [mg/s]) improvement for high NOx values and a 32% (from 1.31 to 0.89[mg/s]) improvement for low NOx values compared to DNN models. Finally, a comparative analysis was conducted against other state-of-the-art predictive models, including LSTM and TCN, and the results show that the proposed method consistently outperforms these models. This superiority was validated through Analysis of Variance (ANOVA) and Tukey’s Honestly Significant Difference (HSD) test.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.