Abstract: Power load forecasting can be divided into long-term prediction, medium-term prediction, short-term prediction and ultra short-term prediction, according to different time horizons of sample sequences. If sample sequences are different, their statistical characteristics are different. Therefore, we should adopt different forecast methods and models in this prediction process. This paper mainly discusses short-term load forecasting. Power system short-term load sample sequences not only have the characteristics of tendency and periodicity, but it also fluctuates greatly, making it difficult to forecast. In the current forecast models, there are some drawbacks, and the accuracy of fitting is not good. In order to improve the accuracy of short-term prediction, this paper established the Optimal Subset Regression (OSR) model for prediction based on the Mean Generating Function (MGF) sequence. MGF adopts (m=INT(n/2)) different cycles to extract sample data from the original sample sequence, and x(n) to calculate their averages to get the new sequence. Then they are repeatedly extended to the length of the original sample sequence and we can get n×m matrix Fn×m called Mean Generating Function matrix. Each column vector fl(0)(t) of Fn×m can be used as a regression predictor of the original sample sequence. Furthermore, in order to fit the high frequency component of the original sample sequence, we make twice-differencing operations of original sample sequences so that we can achieve the effect of high-pass filtering. The first-order difference sequence of x(n) is x(1)(n-1) and the second-order difference sequence is x(2)(n-2). We can get fl(1)(t) and fl(2)(t) by calculating the MGF extended sequence of x(1)(n-1) and x(2)(n-2) respectively, then generating accumulation extended sequence fl(3)(t) by using fl(1)(t), so we can get about MGF extended sequence fl(0)(t), fl(1)(t), fl(2)(t)和fl(3)(t) (l=1,2,..., m). In order to reduce the calculation work, we use all the predictors for roughing, using the sample sequence for simple regression. According to the couple score criterion (CSC), we can get the CSC value of simple regression. Then we make χ2 test and keep qualified predictors, which, with bigger CSC values, are about 1/6 of the original predictors. We take these qualified predictors for free combination, generating 2p-1 subsets, and using these subsets for regression calculation. Then we calculate again the CSC value of the regression result, and select the subset with the maximum CSC value－the optimal subset using as prediction regression equation. According to the requirements of the prediction length, we extend the length of predictive factors of the optimal subset, substituting the extension value into the regression equation, so we can get the forecast results.This paper takes a certain area in northern China as an example and uses hourly load values as the sample sequence, using the above algorithm to forecast four-hour-ahead load values. In this case, we got a total of 45 predictors, then 8 predictors by roughing, and generated 255 subsets. By regression calculation, we determined the optimal subset regression equation. By using the equation, we calculated the fitting value average error at 2.41%. According to the above method extending predictors, we identified the prediction value by regression calculation. The prediction average error is only 4.24%. Compared with principal component analysis of mean generating function model and grey model, the maximum error of prediction model was reduced by 10% or so.