A flood hazard cause classification model for substation flood prevention case text

Flood has brought significant threat to the safe operation of power substations, and it is essential to analyze the flood hazard cause for operation and maintenance measures. The traditional manual analysis is subjective and time-consuming, resulting in a low accuracy and effectiveness. This paper proposes a BERT-based classification model that automatically identifies multiple flood hazard causes. To improve analysis accuracy of hazard cause, we enhance local semantic features related to cause by stacking a TextCNN network on BERT, also extract cause coupling among various cause labels by using a Seq2Seq module. To achieve online application, we reduce model complexity and speed up model inference via a lightweight strategy. Experiments demonstrate that both three strategies and their cooperation are effective. Besides, the proposed model outperforms baseline and state-of-theart models, obtaining the highest score and shortest inference time. Moreover, a case study about real-world substation visualizes the model inference output, assisting engineers to make maintenance decisions for substation flood control and disaster relief. In a word, the proposed model can automatically, and precisely analyze flood hazard cause in power substations, laying a foundation for the online decision assistance system about flood prevention.