Accurate load forecasting is the key to ensure the orderly operation of the power market. In view of the limited environment of single factor forecasting and the ill- considered problems of the split and interaction of variables in the comprehensive model, a generalized additive short- term load forecasting model considering subdivided characteristic variables such as temperature, holidays and load is proposed. Firstly, the correlation coefficients of different factors and loads are subdivided and calculated,and the eigenvalues are selected according to the correlation degree. Secondly, the load fluctuation basis function is determined,and the whole forecasting model is formed according to the subdivided characteristic variables. Finally, the actual operation data of Guangzhou is used to calculate and analyze. The example shows that the generalized additive model based on eigenvalue subdivision is an effective method for short-term load forecasting.