Information Extraction Method of Power Grid Operation Mode Based on Double Layer Labeling Model

In the long-term power grid construction process, the grid dispatching department keeps many adjustment notices of grid operation mode, which record important grid equipment models and their corresponding operating status requirements. In order to facilitate dispatchers to quickly obtain information about abnormal equipment and its operating status when facing with emergencies, it is necessary to study the information extraction method to obtain grid operation rules from such texts. Due to the linguistic characteristics of the entities in the adjustment notices of power grid operation mode such as diverse expressions, uncertain lengths, complex structures and unclear boundaries, a double layer labeling (DLL) model is proposed. By constructing an independent deep learning network at each layer, DLL model first recognizes the outer-layer entities of power grid equipment model and operating state. Then, the key inner-layer entities which make up the power grid equipment model entity was recognized to map the power grid model entity to the corresponding ID. Finally, the unstructured adjustment notices of operation mode can be stored as structured information. The above method is a promising tool to help dispatchers reasonably arrange the operation mode of the power grid and eliminate the hidden dangers of the safe and stable operation of the power grid in time. The DLL model based on two independent Bidirectional Encoder Representation from Transformers and Long Short-Term Memory and Conditional Random Fields networks (BERT-BiLSTM-CRF) is proposed in the article, and the F1 values in the test data set reach 0.932 and 0.963, respectively. The experiment results validate the efficiency of the proposed model for extracting information from the adjustment notices.