Abstract: To tackle the issues of strong concealment of hidden dangers in distribution networks amid highproportion new energy integration in new-type power systems,along with the low accuracy and slow response of traditional single-data perception methods,a multi-source heterogeneous data-driven intelligent perception scheme is proposed.It integrates heterogeneous data like grid operation,equipment status,and environmental meteorology to build a "preprocessing-feature fusion-intelligent recognition"hierarchical framework.By using an improved deep learning algorithm with an attention mechanism,it mines deep-seated data features for precise hazard identification and location.Experiments and engineering validations show 97.2%accuracy,0.958F1score,and 56ms response latency,outperforming traditional