Aiming at the problem of clustering analysis and prediction of residential daily electricity load, a prediction framework based on the fine classification of residential power load patterns was proposed. In order to improve the quality of features used for cluster analysis, feature selection was first implemented based on BIC criteria. Then, the CFSFDP algorithm based on weighted pearson distance is used to realize the accurate identification of the shape of the residential electricity load curve. Next, the LSTM prediction network is improved by a fusion activation function method. Finally, the improved LSTM network is used to predict the finely classified residential power load patterns. The experimental results show that the forecast error index obtained by the method proposed is MAPE= 6. 6792%, which improves the quality of load forecasting and has a good effect in the forecast of residential electricity load.