The load impact of high-horsepower tractors during operation will cause a wide range of fluctuations in the output torque of the engine. In order to reduce the impact of load impact on the tractor power unit, a tractor coupled-split power system configuration with the engine and dual-motor as the power source was proposed to reduce the shift frequency of the power transmission system caused by load impact. A torque allocation strategy was proposed based on Haar wavelet decomposition algorithm and power prediction. Firstly, a priori prediction of power demand for tractor rotary tillage based on radial basis function neural network was researched in which the working load parameters were collected;and then the comprehensive dynamics of tractor loads were mathematically formulated. Then the torque requirements for high and low frequencies were determined by Haar wavelet transform and provided by motor and engine respectively. Finally, the effectiveness and feasibility of the proposed strategy were validated with the hardware-in-loop test. The result indicated that the prediction model of power requirement based on neural network can accurately predict the power demand of driving and power take-off (PTO), and driving end and PTO end root mean square error of predicted values accounted for 7.6% and 7.9% of the maximum power, respectively. The proposed model predictive controller can following tractor torque demand in operation. The torque ripple of the engine was reduced by 35.0% compared with the traditional configuration. And the strategy effectively reduced the torque variation range of the engine and alleviated the adverse effects of excessive shock of the operating load.