Data barriers exist in power metering systems

hindering cross-entity data sharing and integration

which leads to low accuracy in data-driven identification of abnormal electricity consumption behaviors. While federated learning can alleviate data silos

traditional methods struggle to meet the diverse needs of different entities regarding anomaly features. Additionally

issues such as insufficient privacy protection and lack of incentive mechanisms persist. To address these limitations

this study proposes a collaborative optimization federated learning framework that balances privacy and utility. The framework incorporates several key innovations. First

it employs wavelet decomposition to segregate user electricity data into approximation and detail coefficients

separating common and individual characteristics as well as low-sensitivity and high-sensitivity data components. Then

an optimal differential privacy strategy is derived through a master-slave game model

incentivizing power entities to share high-value raw data while balancing privacy protection and data utility. Finally

based on the optimal personalized privacy budget obtained from the game mode

a hierarchical differential protection is applied to highly sensitive personalized models. This approach integrates a novel federated aggregation method

combining average weight parameters from power entities and magnitude weight parameters from metering centers. It enhances the local adaptability of power entity models and the global universality of metering center models while ensuring robust data privacy and security. Experimental results on an abnormal electricity usage detection dataset demonstrate the effectiveness of the proposed framework in improving detection accuracy while maintaining data privacy and utility.