For the delayed effect of the influence of feature on the power load, a two-stage feature selection technique considering delay effect is proposed on the basis of the existing two-stage feature selection technique. Firstly, the optimum prediction reference point between feature and load is searched based delay mutual information (DMI) theory, and the phase space reconstruction is carried out by the optimum prediction reference point. After that, the two-stage feature selection based on DMI is implemented for the above-mentioned phase space, which consist of uncorrelated filtering and redundant filtering. Finally, the filtered phase space is input into the prediction engine to predict. Example is used to analyze lots of variables that may be relevant, which proves that this method can reduce prediction error and time.