It is important to choose the right model according to the trend of the historical data in the process of load forecast model combination. And then, a method is chosen to assign weights according to the features of the models. Even forecast models meet the requirements of the grey correlation degree, the forecast results still have large differences. To solve the question, this paper, according to the feature that the growth rate of load data is non-aftereffect property of Markov chain, and by analyzing the growth rate of load data, uses Markov chain to divide intervals and screens two kinds from the models which have met the accuracy requirement, and adopts the method of variance- covariance to assign weights. Using this method of screening not only can accurately choose the models for combination forecast, but also has a high precision. This work is supported by National Natural Science Foundation of China (No. 51107143).