Neural network for modeling and optimization of internal combustion engines
CHALLENGE - Analyzing performance and quality of combustion in diesel engines for automotive application is a challenging task. Experimental and numerical techniques are very useful to predict the behavior of engines. Optimization based on simulations is impractical due to the number of parameters and the expensive computational cost. Therefore, mathematical models (metamodels or RSms) can be employed to predict performances on the basis of experimental data.
SOLUTION - The neural network algorithm has been applied in order to predict pollutant emissions in a Diesel engine using experimental data. Of a total of 108 points, 89 have been used for training and 19 for validation. The most important task is to obtain a good fit of the training data, avoiding both underfitting and overfitting. Experimental data has been imported in modeFRONTIER where the neural network has been generated, analyzing the influence of the number of neurons of the hidden layer, the mean percentage error allowed in the training set and the maximum percentage error allowed for each point on the model accuracy.
BENEFITS - A neural network metamodel predicting pollutant emissions considering EGR rate, timing and duration of the injectors has been successfully generated in modeFRONTIER. The possibility to change the algorithm parameters enables to avoid both underfitting and overfitting errors while minimizing the computational effort of the training phase. Best performance was obtained with 15 hidden neurons for the NOx levels prediction and with 12 hidden neurons for the THC.