Power prediction tasks are of significant importance in time-series prediction research due to their close association with energy issues. However, when applying traditional clustering algorithms to predict power consumption using time-series data, three main problems often arise. Firstly, clustering results may be of poor quality due to insufficient users in the cluster. Secondly, users in the same cluster may have similar total power consumption values for the year, but their data may differ in some fine-grained time periods. Finally, users in different clusters may still have similar data in certain time periods. To address these issues, we propose a new Two-Stage Clustering Framework (TSCF). Our framework can divide user data into proper data segments, making it applicable even if there is only one single user, thereby addressing the issue of weak clustering performance caused by insufficient users. Additionally, TSCF segments data into finer pieces to address the issue of fine-grained dissimilarity between users in the same cluster, and groups similar data segments into the same cluster to address the issue of similar data in different clusters. Finally, TSCF leverages the model selection process to find the model that best fits current user data. Extensive experiments are conducted on real-world power consumption data of commercial users in Taiwan and the United States to compare our proposed framework with various baseline approaches. The results show that our proposed framework outperforms baseline methods by approximately 49.67% and 144.22% on the MAE and MAPE indicators, respectively.