In response to wildfires that threaten the stable operation of distribution lines, it is important to establish a fire risk prediction model for distribution lines. However, the scarcity of data on wildfires causes sample imbalance, affecting the accuracy of the model. To this end, based on the influencing factors such as meteorological, geographic, combustible, and social factors, support vector machines and the idea of cost sensitivity are used to assign more weight to minority samples. Recursive feature elimination is used to select features that favor minority class classification. On this basis, a fire risk prediction model for distribution lines based on the random forest-adaptive boosting algorithm (RF?AdaBoost) is constructed. Finally, a 10 kV line corridor area in Xichang City, Sichuan Province, is selected to carry out an example verification. Ten-fold cross validation is used and compared with other algorithms. The results show that the recall rate of the method in this paper increases to 76.67%, which lessens the impact of sample imbalance on the model performance, reduces the misclassification of wildfires, and provides a basis for wildfire prevention and control in line corridors.