Abstract: This project is based on a sliding time window ensemble learning algorithm, which deeply mines and analyzes the daily electricity consumption data of over 400000 social assistance recipients in Zhejiang province, and models them separately according to different cities, achieving active identification of disadvantaged groups and allowing more disadvantaged groups to fully enjoy the benefits. Based on the user tags provided by the Civil affairs department, construct different assistance scenarios, and apply multi scenario temporal anomaly detection algorithms to implement classification management and early warning mechanisms for different types of disadvantaged groups. This will provide personalized electricity care for the government to improve assistance work, ensure special electricity use, and safeguard life safety by leveraging innovative data application value.