Find in Library

Aimed at improving the prediction accuracy of engine emissions under driving conditions which are not covered by the training set, an extrapolated Gaussian process regression (GPR) method is proposed. First, the training set data is fed to the GPR model for pre-training, and then a wide-area input set is constructed by uniform sampling between plus/minus three standard deviations, and the hyperparameters are optimized for the goal of minimizing the prediction variance of the input set. The test results on a direct injection gasoline engine show that the mean absolute error of emission prediction using the extrapolated GPR is 0.53411, which is 24.27% lower than that of using the traditional GPR and 36.32% lower than that of using the back propagation (BP) neural network, which signifies the effectiveness of the proposed method in terms of reducing test costs and improving the accuracy of emission prediction during real driving.