Neural network prediction of cycle-to-cycle power variability in a spark-ignited internal combustion engine

Abstract

Cycle-to-cycle variation (CCV) limits how lean a spark-ignited (SI) internal combustion engine (ICE) can stably operate at, restricts efficiency, and increases emissions through incomplete combustion. Therefore, a way to cleaner, more efficient SI ICEs is to minimize the CCV. Current methods to study CCV include experimental investigations and CFD-based numerical simulations. This study, in contrast, investigates the ability of neural networks to accurately model the indicated mean effective pressure (IMEP) and its coefficient of variation (COV of IMEP). Experimental data from a previous study of spark-ignited propane/air combustion in the TCC-III engine was used to train and evaluate a neural network. An optimized network was generated that utilizes 109 experimental inputs and is operated with 15 neurons in one hidden layer to determine IMEP for 18 engine operating conditions, with 625 individual consecutive engine cycles for each condition. The impact of training set size and the number of input parameters was also investigated. The average deviation for IMEP from the experimental measurements is 0.7–2.2% for the training data set and less than 12% for the entire predicted range of operating conditions. Data sets consisted of tests under rich, lean, and stoichiometric conditions without and with 9% nitrogen dilution. Predicted COV of IMEP strongly correlates with experimental data (R2 = 0.8453). However, a systematic over prediction of COV of IMEP for low COVs was observed while higher COVs were under-predicted by the neural network. The cause for this systematic behavior has not yet been identified but histograms of the predicted IMEP data indicate that this could be related to missing physical parameters that have a significant impact on combustion variability.