Accurately predicting the actual working condition pollutant emissions of tractors and other diesel machinery is of great significance in the establishment of emission inventories and regional pollutant emission control. Based on the measured data such as engine speed, fuel consumption, combustion ratio, CO, HC, NOX and PM under different operating conditions of the tractor as the data source, a Deep Extreme Learning Machine (DELM) prediction model was established, and the tractor idling speed, and the pollutant emissions under basic working conditions such as tractor idling, walking and rotary tillage are predicted. In order to further evaluate the adaptability of the DELM prediction model, it is compared and analyzed with support vector machine (SVM) and back propagation neural network (BPNN) models. The results show that 1) The DELM model has certain advantages in predicting the emission time series. It predicts that the average root-mean-square error of the NOX, HC, CO, and PM emissions of the tractor in the three states of idling, walking and rotary tillage are 5.269×10-5, 5.195×10-5, 5.135×10-5 and 2.795×10-5. 2) The DELM model is compared with SVM and BP and it is found that the DELM model has significant advantages in robustness and adaptability. 3) The high accuracy and generalization of the DELM method provide ideas and methods for predicting mobile source exhaust emissions based on engine state data.