The prediction of diesel exhaust emissions is of great importance for environmental regulations. However, traditional methods are limited in prediction accuracy because they ignore the feature redundancy and nonlinear coupling present in On-Board Diagnostics (OBD) data. This paper finds that feature variables are the core of emission prediction. Therefore, this paper proposes a hybrid prediction model RF-MLP that integrates random forest and multilayer perceptron. The model extracts key features by constructing a feature screening mechanism, removes interfering noisy features, and then achieves high-precision fitting of complex nonlinear relationships of emissions through a network of multilayer perceptron models. The experiments of this model on the Hefei diesel vehicle emission data set show that the MAE and MSE metrics of this method are reduced by 35.6% and 73.5%, respectively, compared with the optimal model decision tree algorithm, which provides a solution with high prediction accuracy for diesel vehicle emission prediction.