Prediction of Diesel Engine Emission Characteristics Based on GRU Neural Network

To improve the accuracy of diesel engine emission predictions and reduce experimental costs, a method based on a gated recurrent unit (GRU) network was proposed. First, the importance of input variables was analyzed using the random forest algorithm, and the input variables for each emission parameter were determined. Then, the hyperparameters of the GRU prediction model were automatically optimized using the particle swarm optimization algorithm, and the model was trained and validated using the optimal hyperparameters. In the training set, the coefficients of determination (R²) for NO_x, CO, CO₂, THC, exhaust temperature, and exhaust pressure are 0.985, 0.994, 0.991, 0.986, 0.979, and 0.998, respectively. The corresponding values in the validation set are 0.962, 0.969, 0.987, 0.954, 0.975, and 0.992, respectively. Finally, the generalization performance of the GRU prediction model was tested, and the model was compared with other neural network prediction models. The results show that the average R² value is 0.930 in the test set, which is 5.9% and 4.3% lower than the average values for the training and validation sets, respectively, indicating that the GRU prediction model has excellent generalization ability. Compared with other prediction models, the GRU model also demonstrates superior predictive performance.