Household short-term load forecasting is one of the essential work of electricity utilities, which is of great significance to the construction of smart grids and the safe operation of power systems. As is the basic support of supply and demand balance, household load forecasting is expected for higher accuracy. However, the high fluctuation in the household load makes the performance of the existing prediction model limited. Considering that the stochastic characteristics of household load forecasting, a new aggregated household load forecasting framework based on clustering method and multi-extreme gradient boosting (MXG-Boost) is proposed. K-means algorithm is utilized to divide the daily load (DLCs) of a targeted resident into several clusters so that the high fluctuation can be classified. For each cluster, a XGBoost model is built to learn the load. Then the DLCs of each cluster are applied to training the corresponding XGBoost model. The MXGBoost models are aggregated to forecast the records in the next hour with prior weights. The results of case studies show that the MXGBoost model based on clustering method can obtain good prediction performance.