Abstract: Agricultural machinery is one of the important components of modern agriculture. In recent years, the number of agricultural machinery has continually increased, as well as the fuel consumption caused by agricultural machinery. The fuel consumption of agricultural machinery is directly related to agricultural production costs and the vital interests of farmers. Estimating the fuel consumption of agricultural machinery is of great significance in environmental governance, agricultural machinery operator evaluation, and agricultural cost input. Different from road vehicles, the factors affecting agricultural machinery fuel consumption seem to be more complex. Taking driving conditions for the only consideration cannot accurately predict the fuel consumption of agricultural machineries. Random forest, as a typical representative of ensemble learning, has many applications in various fields and has strong fitting ability for nonlinear data. It is widely used in the research of vehicle fuel consumption prediction. The purpose of this article is to realize the fuel consumption prediction of the grain harvester, WORLD 4LB-150AA, during working in the farmland. Based on the engine operating condition data and driving condition data collected by the harvester CAN(Controller Area Network) bus and GPS (Global Positioning System) terminal, seven indicators are constructed, including engine mean torque, engine mean speed, average speed, mean acceleration, mean deceleration, acceleration standard deviation and deceleration standard deviation. The acquisition frequency of the average data is 1.3 s, and the total number of the records is 130 788. Agricultural machineries that provided the data worked in six provinces including Liaoning, Jilin Province, Shandong, Jiangsu, Zhejiang, and Hubei. At the same time, by calculating the Spearman correlation coefficient, the correlations between seven indicators and fuel consumption were explored. According to China's agricultural divisions, the six provinces are divided into three regions: northeast region, plain region, and hilly region. Then, the fuel consumptions of the same grain harvesters in different regions were analyzed. Above the analysis, the fuel consumption prediction model of the harvester based on Random Forest was established, and compared with the one based on support vector machine. The results showed that the fuel consumption is correlated with all indicators. Among them, the fuel consumption is highly correlated to engine mean torque, engine mean speed and average speed, all with the correlation coefficient above 0.6, followed by mean acceleration, mean deceleration, acceleration standard deviation and deceleration standard deviation, whose correlation coefficients are above 0.4. There are significant differences in the fuel consumption of harvesters working in different regions. Among them, areas with high output per unit area are also relatively high in fuel consumption. Moreover, the Random Forest based model can realize the higher accurate prediction of fuel consumption during harvester operation. The root mean square error RMSE is 0.14, the average absolute error MAE is 0.24, and the coefficient of determination R2 is 0.84. The proposed method in this paper can provide a reference for the optimization of working conditions of agricultural machinery and precise fuel consumption monitoring.