Comparative research on DNN and LSTM algorithms for soot emission prediction under transient conditions in a diesel engine

초록

Deep learning approaches were applied to predict soot emissions under transient conditions in a diesel engine using the worldwide harmonized light vehicles test procedure (WLTP) cycles. The accuracies of deep neural networks (DNN) and long short-term memory (LSTM) models were compared to predict emissions. The accuracy of the LSTM model had an R-2 value of 0.9761, which was higher than that of the DNN model, with an R-2 value of 0.9215. The mean absolute errors (MAEs) of the WLTP cycles predicted by the LSTM model were between 0.30 %-1.47 % compared with the maximum measured values in WLTP cycles. For local prediction, the LSTM model followed the fluctuations in the data and local peak values well. However, the calculation time of the LSTM model is longer than that of the DNN model. Researchers should consider the purpose of the model based on the accuracy-time trade-off relationship between DNN and LSTM models.

키워드