Short-term load forecasting is the basis of safe and economical operation of power grids. The accuracy of selecting similar days directly influences the degree of accuracy of the short-term load forecasting algorithm. After an in-depth analysis of the common laws of such factors as the meteorological factors and the day type that impact power loads， a new algorithm of selecting similar days for the short-term load forecast is proposed. The algorithm can distinguish between the key factors in many cases. The simulation results demonstrate that it can select the most suitable similar day from a large number of historical data to improve the accuracy of short-term load forecasting with relatively strong adaptability.