Abstract: The stable operation of the power grid is crucial for the efficient function of modern society’s production and daily activities. Therefore, it is imperative to monitor and analyze power outages at various levels of the system. However, the implementation of power outage monitoring through the addition of edge nodes can incur significant costs in terms of construction and maintenance. To address this challenge, in this paper we propose to leverage the data from network devices such as smart homes outside the power grid to enable real-time, high-precision, and low-cost power outage fault perception with broad coverage. Currently, the deep learning-based power outage models necessitate parameter retraining for different scenarios, lack lightweight deployment, and exhibit limited real-time performance. In view of this, we introduce an end-to-end lightweight power grid outage perception model, which utilizes multi-source heterogeneous data fusion method. Specifically, the graph clustering algorithm is employed to extract the characteristics of power outage events of user clusters, thereby obtaining a power outage prior as a prompt for further research and judgment on power outage. Moreover, adaptive prompt learning is introduced to achieve lightweight reuse of the same model parameters, making power outage research and judgment possible in different scenario levels. Finally, the group sparse algorithm is applied to optimize the computational efficiency of the outage model. This approach has significant practical value and promising application prospect, and contributes to improving the reliability and stability of the power system.