The increasing intrusion of extreme weather on power system has brought challenges to the stable operation of power system. Traditional fault warning methods for power grids have limited granularity and cannot effectively solve the imbalance of training samples. Therefore, a distribution network fault warning evaluation method based on adaptive regional ensemble learning is proposed in this paper. First, the collected static and dynamic multi-source heterogeneous data are preprocessed. Then, the upper-level learning algorithm realizes adaptive regional division through the association between the distribution network area and the distribution network data characteristics, and the lower-level learning algorithm trains the model for the samples collected in the area based on the improved cross entropy loss. Finally, the verification set accuracy method is used to determine the distribution network fault warning level. The effectiveness of the evaluation method proposed is verified through an actual data from a province. By comparing with existing methods, it shows that the adaptive regional ensemble learning strategy has obvious advantages in the fine-grained distribution network fault warning task.