A comparative investigation of advanced machine learning methods for predicting transient emission characteristic of diesel engine

• Published in: Fuel (Guildford) 2023-10, Vol.350, p.128767, Article 128767
• Main Authors: Liao, Jianxiong, Hu, Jie, Yan, Fuwu, Chen, Peng, Zhu, Lei, Zhou, Quan, Xu, Hongming, Li, Ji
• Format: Article
• Language: English
• Subjects: Diesel engine; Emission characteristic prediction; Hyperparameter optimization; Machine learning; Variable importance analysis
• ISSN: 0016-2361; 1873-7153
• DOI: 10.1016/j.fuel.2023.128767

[Display omitted] •Seven advanced machine learning methods applied to the transient emission characteristic prediction of diesel engine were introduced and compared.•The correlation between input parameters and emission characteristic parameters was analyzed based on Pearson and Spearman correlation analysis methods.•The optimal hyperparameters for each machine learning method were obtained using GA and PSO algorithms.•A hybrid prediction model that combines multiple suitable algorithms was proposed to have excellent performance in all emission characteristic prediction. Machine learning method provides a promising way to predict the transient emission characteristic of diesel engine due to its many advantages such as short computation time, low consuming, high prediction accuracy and good robustness. In this paper, seven machine learning methods were discussed in detail and compared with each other, including ANN, SVM, NARX, LSTM, GRU, Transformer and TCN, to find the most suitable machine learning method applied to the transient emission characteristic prediction of diesel engine. Each machine learning model was trained, validated, and tested based on WHTC and WHSC cycles and the R2, MAE and RMSE were used as evaluation metrics. In addition, the correlation between input parameters and emission characteristic parameters was analyzed based on Pearson and Spearman correlation analysis methods, and the top six important parameters were considered as model inputs. The hyperparameters of each model were optimized by GA and PSO algorithms to identify optimal model structure. The results showed that there was no machine learning method that can show excellent overall performance in all emission characteristic prediction. In the NOx prediction, GRU and TCN models performed the best overall performance. The optimal method for CO and CO2 was TCN and LSTM, respectively. Transformer model had relatively better overall performance in the THC prediction although its prediction curve showed oscillation. If the system configured for the machine learning model has sufficient computing power, more complex NARX or GRU models are recommended for exhaust temperature prediction. Otherwise, TCN model is chosen if the computing power is relatively limited while some prediction performance is sacrificed. In particular, SVM model with simplest structure showed best overall performance in the exhaust pressure prediction. Finally, inspired by Ensemble Learning methods, a hybrid pre