The multi-resolution SVM（M-SVM） is adopted for the regressive estimation of the short-term load forecasting curve after a study of the approaching and extending capabilities of SVM， in which several kernels of different scales of SVM can be used simultaneously to approximate to the target function and improve the effectiveness of generalization and approximation in the local area model. Additionally， the data set is arranged into 96 regression functions for every time spot in a whole day， respectively. The performance of the model proposed is evaluated through a comparison with other algorithms such as the traditional SVM and artificial neural network（ANN） methods. Test results of an actual power system show that it has better local approximation and generalization capabilities when appropriate numbers and parameters of the kernels of SVM are chosen.