The summer load is greatly affected by meteorological factors such as temperature, showing the characteristics of strong randomness and large fluctuation. To solve the problem of the low short-term load forecasting precision of the existing models in Summer, it is proposed to decompose the temperature into daily periodic components and fluctuation components while the load components are decomposed, which is beneficial to accurately grasp the impact of short-term weather fluctuations on short-term load forecasting. After analyzing the variation feature of each load component and its impact on the forecasting accuracy of the overall load, the forecasting method for each load components is designed respectively according to their different feature. Taking the temperature fluctuation components into consideration while forecasting the weather-sensitive load, a short-term load forecasting model is constructed based on the XGBoost algorithm. The historic summer load of a middle city of China is chosen to establish the training samples, the results of 96 time points load in August 2017 show the proposed forecasting model and algorithm are effective.