Daily load forecasting is a basic role of power market. Most of load forecasting methods use one same model in one day, regardless of the change of load composing and characteristic at different time segments. A new segmented multi-model combining load forecasting strategy was proposed in this paper. According to different load composing and characteristic, 96 points daily load was separated into many time segments. At each time segment, a multi-model combining load forecasting, composed by multivariate linear regression, grey prediction, SVM and neural network forecasting, was used to forecast load. The forecasting results of a city in east China showed that, the MSE forecasting error of 96 points daily load is only about 1.78%. The method can satisfy the request of real power system well.