The new power system imposed higher requirements on the safety

real-time performance

and effectiveness of load regulation to achieve "precise matching and balanced coordination". However

the existing load regulation architecture and aggregators' day-ahead coarse regulation strategies can not fully utilize the real-time data from smart meters to accomplish safe and precise regulation

making it difficult to harness the flexibility of user resources and limiting the regulation effect. To address this

this paper proposed a residential load safety regulation architecture and a regulation strategy based on the smart home platform. For the first time

it employs a converged dual-network indoor communication method and utilizes neural network weight encoding in information dissemination

which can effectively isolate malicious network attacks and avoid electricity meter data leakage. Simultaneously

this architecture significantly enhances data interaction efficiency and response speed by leveraging real-time perception from internet of things (IoT) electricity meters and coordinated optimization of load regulation with the smart home system. Finally

this paper proposes a Q-value mixing (QMIX) algorithm generated by the optimal control strategy based on a partially decoupled deployment with joint training model. The agents adopt a dual-model joint training and distributed execution approach

enabling IoT electricity meters to rapidly perceive changes in indoor electrical quantities and achieving load safety regulation through one-way encrypted broadcast of meter data. Simulation results demonstrate that the proposed strategy effectively resolves the conflict between user privacy protection and real-time regulation performance

providing a novel technical approach for safe and precise load regulation in the new power system.