The medium and long term load forecasting is the basis of power planning, investment, production, scheduling and trade, which plays an important role in electric power safety and economic operation. In China, it has the increasing uncertainty and the uncertainty of random variation to forecast the medium and long term load. Thus we can regard it as a typical grey system. However, the traditional grey prediction method cannot be adapt to the needs of the load forecasting gradually. It need to be rich and perfect with the continuous improvement of power system complexity and power marketization degree. This paper studied the modelling mechanism of grey prediction model. Then we analyzed the problems existing in the model, including the boundary value problem, the background value structure problem and the least squares parameter identification problem. This paper put forward an optimization method to directly identify the boundary value x(0)(1), the developing coefficient a and grey coefficient b using ant colony algorithm according to the time response expression of GM(1,1) model, so that it established an optimized GM(1,1) prediction model based on ant colony algorithm. This model can fix the impact of boundary value, and also avoid the errors brought by the background value construction and the least squares parameter estimation. It can verify the effectiveness of the proposed optimization model through the load data simulation. And it can improve the prediction accuracy effectively.