Abstract: Short-term household power load forecasting has played an increasing important role in smart grid. To further improve the accuracy of the forecasting, a state frequency memory network-based short-term household power load forecasting model was proposed. Firstly, the k-means clustering method was utilized to classify the families possessing the same electricity consumption mode into the same category. Secondly, the wavelet denoising technology was applied to the load data. Finally, a state frequency memory network model was constructed to perform batch of household power load forecasting. In the proposed model, the discrete Fourier transform was led in to decompose the memory state into multi frequency components, and by means of the combination of these frequency components the future electricity consumption was forecasted. The mean square error (abbr. MSE) , root mean square error (abbr. RMSE) and mean absolute error (abbr. MAE) were used to evaluate the proposed model. Taking the load forecasting of the next day for example, comparing the results obtained by LSTM, which behaves the best in this field, with those obtained by the proposed model, the error of forecasted results of three kinds of household power load has reduced by 21.6%, 11.4% and 15.4% respectively, thus, the effectiveness of the proposed model is fully verified.