A comprehensive evaluation of neural network (NN) methods to model exhaust gas aftertreatment reactors is presented in this study. A combination of non-linear autoregressive model with exogenous inputs (NARX) network and static NN is utilized to capture the dynamic behavior of the state and outlet variables, such as coverage fractions, temperatures, and conversions, of monolith reactors used in the abatement of automobile emissions. The efficacy of this methodology is illustrated by modeling the diesel oxidation catalyst (DOC) and selective catalytic reduction (SCR) reactors of a diesel engine exhaust gas aftertreatment system. The final NN topology is designed by carefully comparing the model performances in the training and testing datasets. Model parsimony quantified by Akaike’s Information Criterion is employed for choosing the number of input and feedback delays in the NARX network. The novelty of this work lies in generating quality training and validation data using a physics-based system level model comprising of the engine and aftertreatment reactor components. The accuracy and generalizability of the NN models are demonstrated by comparing the model predictions with physical model simulation results in five disparate test datasets used in vehicle certification protocols. Overall, the NN models are found to reproduce all the qualitative trends in the various state and output variables while needing minimum computing resources resulting in better than real-time performance. Therefore, with appropriate training, these models can be embedded in the electronic control unit (ECU) and utilized for model-based predictive control systems.
Read full abstract