Lightning is one of the main causes of voltage sags in power grid. Accurate estimation of the severity of voltage sags caused by lightning can provide a basis for developing optimal management plans and siting sensitive users. In this paper,a data-driven self-learning estimation method for the severity of voltage sags is proposed. Firstly,based on the mechanism of voltage sags caused by lightning,the parameters involved in mining are selected by the monitoring information in lightning location system and power quality monitoring system. Secondly,the influence of discretization results on the accuracy of rules is reduced,and the number of discretization intervals for different parameters is determined by using discretization evaluation indexes. Then,to solve the problem of low efficiency of mining algorithm when the grid database changes dynamically,the incremental learning-based association rule mining algorithm to continuously update the mined rules,which gives it the ability of self-learning. Finally,a weighted Euclidean distance based on the integrated assignment method is proposed to evaluate the severity of voltage sags in real scenarios. The results of the empirical analysis by monitoring data of a regional power grid and simulation data of IEEE 30-node prove that the method in this paper can accurately mine valuable rules in reality and realize the severity assessment of voltage sags of the concerned nodes.