The load forecasting is the important basis of the energy management systems for the evaluation and diagnosis, optimized control, and scheduling of the energy subsystems in buildings. In order to obtain real-time and high accuracy load information, this paper proposes a short-term load forecasting method based on data mining for buildings. It firstly finds the sample datasets that are similar to the forecasted time points from the historical data, and then performs the K-means cluster analysis on the meteorological data, such as temperature, humidity, barometric pressure, etc., and finally adopts the support vector machine (SVM) for short term forecasting. The practical application results prove that the eMAPE of the proposed method is 1.34%, and the probability of the relative error less than 1% is 67.5%, which are obviously better than that of the ARIMA, SVM and DMSVM without meteorological data.