Selective Catalytic Reduction (SCR) systems are crucial for automotive emissions control, as they are essential to comply with stringent emissions regulations. Model-based SCR controls are used to minimize NOx emissions in a broad range of real-word driving scenarios, constantly adapting the urea injection to diverse load and temperature operating conditions, also accounting for different catalyst ageing status. In this framework, Neural Networks (NN) based models offer a promising alternative to reduced-order physical models or map-based controls.
This study introduces a hybrid modeling approach for SCR systems, leveraging the integration of machine learning techniques with detailed physics-based models. A high fidelity 1D-CFD plant model of a SCR catalyst, previously calibrated on experimental data, was used as digital twin of the real component. A standardized simulation protocol was defined to virtually characterize the SCR thermal and chemical behavior under the full range of operating conditions typically covered during the real operation of the system. The generated dataset, including hard-to-measure physical quantities such as the catalyst wall temperature and the ammonia storage, was used to train and validate the neural network models. In particular, Recurrent Neural Networks (RNN) were used to mimic catalyst wall temperature and ammonia storage temporal trends, while Feed Forward Neural Networks (FFNN) were applied to model SCR outlet temperature and species concentrations such as NOx and NH3.
The NN-based SCR model predictive capabilities were assessed against experimental driving cycles data. The results show that the NN model is able to accurately capture the non-linear characteristics of the system behavior, even under the highly transient conditions typical of real-driving scenarios, thus confirming the reliability of the proposed methodology.