Artificial Neural Network Based Prediction of Engine Combustion and Emissions from a High Resolution Dataset

Abstract

Development of new advanced fuels require more efficient methods to reduce costs. Artificial neural networks can be used in the fuel designing process, but the dataset creation can be expensive. This paper aims to create highly accurate multilayer perceptron type artificial neural network models to predict a medium duty commercial diesel engine's combustion and emission properties. A high-resolution dataset with 6277 samples was used for the training, and the resulted models will be used for future researches on cost optimization. The NOx and PM emission, peak combustion temperature, peak pressure rise rate, indicated mean effective pressure, start of combustion, duration of combustion, ignition delay, brake specific fuel consumption and brake thermal efficiency was predicted from the engine speed, torque and high-pressure exhaust gas recirculation valve position. First, the cost-efficient method of high resolution dataset creation is described, then the results of the predictive models are presented. The mean squared error for the scaled dataset, and the root-mean-square error, mean average percentage error, correlation coefficient and determination coefficient for the unscaled dataset was used to evaluate the performance of the resulted models. In addition the most informative prediction error plots are also presented. It was found that the high-resolution dataset resulted really accurate models that can be used for continuing the cost optimization research.