In Advanced Metering Infrastructure (AMI), In order to minimize data transmission and data simplification of smart meters, a lightweight framework based on prediction is proposed. Firstly, a decision tree is established to find the relationship between the forecasting method and the statistical characteristics of electricity consumption data. Then, the time series of electricity consumption data is analyzed to extract the statistical characteristics. Finally, in order to increase the adaptive ability of the framework to the changing data modes of smart meters, a supervised learning scheme is adopted to switch to the forecasting method which is most suitable for the current data modes in real time. Ten users" electricity consumption data in one year are extracted from the data set. The total number of records collected by each user is 17600. The experimental results show that the proposed framework can achieve high data simplification accuracy (DRA) and data simplification rate (DRP), with the highest DRA of 96.7% and DRP of 98.6%.