Improving prediction accuracy of ignition model by weighting using machine learning

Abstract

Compression auto-ignition engines have a high potential to achieve both high thermal efficiency and low emission of harmful exhaust gases, but they have a problem of difficult control of ignition and combustion. To realize such control, model-based control has attracted much attention, which requires an ignition model with high prediction accuracy and light computational load. In this study, we developed a machine learning method to enhance the prediction accuracy of an existing empirical model of ignition delay period. The method has multiple existing empirical models, each with different model parameter values, and a neural network that weights the prediction values of them considering the driving conditions and in-cylinder gas states. To obtain data for training and evaluation, transient operation experiments of a diesel engine were conducted under conditions that relatively premixed combustion occurred. The neural network and the model parameters were trained by the stochastic gradient descendant algorithm using the training data, and the proposed method showed much better prediction accuracy for the test data than the single empirical ignition delay period model adapted using the training data. Investigation on each ignition delay period model of the proposed method after training revealed that the prediction values of each model are highly correlated, indicating that the proposed method does not take advantage of the availability of multiple models. However, the proposed method successfully realized the state that multiple ignition delay period models make different predictions, and the neural network weights them based on the driving conditions and in-cylinder gas states.

Keywords

• internal combustion engine
• ignition model
• model-based control
• machine learning
• neural network