A hybrid model for predicting transient emission characteristics of diesel engine based on multiple machine learning algorithms

PEMS (Portable Emission Measurement System) is a high-precision device for measuring vehicle emissions, but its maintenance cost is relatively high. To address this issue, this study constructs specific models for predicting CO2, THC, CO, and NOx emissions by using transient data from PEMS and On-Board Diagnostics (OBD) collected through drum tests. By integrating the engine operation mechanism with machine learning methods, the study analyzes the correlation of operating condition parameters and selects input features for the models. Through comparative analysis of the performance of three algorithms—BPNN (Backpropagation Neural Network), Random Forest, and SVR (Support Vector Regression)—specific models for each emission type are developed. The results show that the overall prediction error of the specific models is less than 10%, with errors for CO2, THC, and NOx below 5%, demonstrating significant advantages in predicting transient emissions of diesel engines. Based on their high-precision characteristics, these models have the potential to detect and diagnose gas emission sensors, replacing faulty sensors to ensure the normal progress of experiments. Meanwhile, using the concept of 60-s cycle segment error and setting a 10% threshold to calculate the compliance rate, this method provides a new effective approach for evaluating a model’s prediction accuracy of data segments, offering a novel means for precision assessment of prediction models.