Development of an LSTM-CCF-MA Model for Predicting NOx Emission and Exhaust Temperature of a Diesel Engine

The continuous reduction of NOx emission limits for heavy-duty diesel engines poses challenges to the control strategies of diesel engine. Developing an accurate prediction model for NOx emission and exhaust temperature is of great importance in reducing NOx emission. In this study, a time-series prediction model for exhaust temperature and NOx emission is built using LSTM neural network. The model's inputs are determined using sensitivity analysis. It can be also found that the prediction of exhaust temperature is highly sensitive to the past values of exhaust temperature and NOx emission. The impact of different types of cost functions on the model is investigated. According to the predictive abilities (the average after ten training runs) of models using different cost functions, a combination of the Mean Absolute Error (MAE) cost function and Huber cost function is selected to further improve the model performance. By introducing a novel combination cost function, multi-head attention mechanism, and convolutional neural network approach into the LSTM model, the LSTM-CCF-MA model was found to yield the best prediction results for NOx and exhaust temperature. The goodness of fit for all the training and test datasets exceeded 0.97.