Text Classification Method for Analysing Accidents in Power System Operation

In order to ensure the safety of equipment and personnel in power system operation, power utilities make a lot of efforts to prevent the occurrence of accidents. It is very beneficial to investigate and learn from the past accident and near miss narratives. As the number of text records related to accidents keeps increasing every year, it will be difficult and time-consuming to analyse these data by manual. In this paper, an improved BiLSTM neural network based method is proposed to realize the automatic classification of the causes behind the accidents. The proposed model consists of embedding layer, encoder and decoder, which are further improved by introduced the mechanism of attention, dropout, layer normalization and residual connection. Comparison studies are conducted on different models, including textCNN, BiLSTM and BiLSTM-based improved models. The results proved the effectiveness and superiority of the proposed model.