Neural Network-Based Prediction of NH₃ Leakage from SCR Systems for Diesel Engines

In selective catalytic reduction (SCR) systems, the urea injection control strategy is central to the control of NO_x emissions. When urea is over-injected, ammonia leakage will occur downstream of the SCR. A neural network-based NH₃ leakage prediction model for diesel engine SCR systems is proposed in order that the dosing control unit (DCU) can reduce the corresponding urea injection volume according to the NH₃ leakage when calculating the urea injection volume. Back propagation (BP) neural network model and gated recurrent unit (GRU) model are developed respectively by code compilation software to predict the NH₃ leakage. The genetic algorithm (GA) is used to find the optimal parameters of the two different models. Bench tests are conducted to evaluate the model accuracy. Under historical test data, the root mean square errors of the final GA-BP and GA-GRU models are 3.142 ppm and 2.378 ppm, respectively. The percentage of cumulative NH₃ leakage prediction error of GA-BP and GA-GRU are 4.808% and 3.745%, respectively. The results show that the method of using neural network for NH₃ leakage prediction is feasible, and GA-GRU model is better than GA-BP model in predicting NH₃ leakage. This provides the basis for developing DCU to reduce NH₃ leakage.