In order to improve the prediction accuracy of agricultural machinery total power then to provide the basis for the agricultural mechanization development goals, the paper used gray GM(2,1) model in the prediction. Through the introduction of parameter λ to correct the background value and parameter ρ for multiple transformation on the initial data, the model was expanded to GM(2,1,λ,ρ) model and prediction accuracy was improved. Because of the nonlinear traits between parameter λ,ρ and the prediction errors, they are difficult to be solved. The paper used Particle Swarm Optimization (PSO) to search the best parameter λ,ρ, then combination forecast model of PSO-GM(2,1, λ,ρ) was constructed. In order to avoid incorrect selection of inertia weight w causing the global search and local search imbalance, the paper used Decreasing Inertia Weight Particle Swarm Optimization, in which parameter w gradually decreases from 1.4 to 0.35. And agricultural machinery total power was predicted based on Zhejiang province’s statistics. Predicted results show that the combination forecast model prediction accuracy is higher than the gray GM(1,1) model and the model better fits the data. The forecast of the agricultural machinery total power of this combination forecast model is feasible and effective, and should be feasible in other areas of agriculture prediction.