The influence of temperature on the load has a cumulative effect, which leads to a big difference on the load even the weather conditions and the day type of the two days are the same. The current similarity data selection methods which take weather conditions and the day type as the characteristic parameters cannot guarantee the accuracy of prediction. Thereby, a dynamic similar method is proposed, and a new short-term load forecasting method is proposed by combining it with the existing methods. The daily average load and the load curve of candidate prediction day are forecasted respectively by using decoupling model. The daily average load is predicted by the dynamic similar method and the static similar method is used for the load curve prediction. In the case study, the first day of holiday and the day of continuous high temperature are taken for the load forecast as an example. Through instance simulation and calculation of a district in Beijing, it is shown that the proposed method can improve the accuracy of the holiday load forecast. This work is supported by State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources (North China Electric Power University) (No. LAPS16003).