An efficient combinatorial modeling to forecast short-term power load is proposed. As the performance of individual models varies in different circumstances，the weights of the combinatorial forecast model should vary with the variation of circumstances. Here the historical power load is classified as three parts，the training set，validation set and test set and four models，namely，the autoregressive moving average（ARMA），generalized autoregressive conditional heteroscedasticity（GRACH），artificial neural network（ANN）and support vector machine（SVM），are selected as candidate models. For short-term load forecast，the circumstance of the coming day is compared with those of past hours，past days，past months and past seasons，and then clustered into the same category by fuzzy Hopfield neural network（FHNN）. The combined weights are obtained according to mean absolute percentage errors of different models. Then the combinatorial forecast model with ARMA-GRACH-ANN-SVM weighted by FHNN clustering is obtained. A case study shows that the proposed combinatorial model outperforms other forecast methods.