Research On Grid Fault Event Extraction And Emergency Plan Recommendation Method Based On BERT

In order to improve the efficiency of repairing grid faults and reduce economic losses caused by grid faults,Intelligent grid emergency plan recommendations have gradually become the focus of attention in the field of power grids.The key of emergency plan recommendation lies in text recommendation and construction of emergency plan knowledge base.The construction of emergency plan knowledge base requires the support of event extraction technology.The traditional pipeline event extraction method has the problem of transmission error and the extraction effect is not good.there are a lot of proprietary words in the text of power grid faults,and the existing text recommendation methods are difficult to apply directly to the field of power grids.Therefore,this thesis is devoted to the research of grid fault event extraction and grid emergency plan recommendation method.The main research works of this thesis are as follows:Firstly,Aiming at the problem of grid fault event extraction,this thesis proposes a method based on BERT embedding to extract the fault events in the grid fault case text.First,pre-process the text according to the characteristics of the grid fault case text,including special word replacement and de-stop words,to reduce the impact of characteristic information loss and noise,and make the text more suitable for extraction tasks;A BERT pre-trained language model is used to generate sentence vector expressions,then Bi-directional Long Short-Term Memory networks are used to capture text features.There are two branches behind the coding layer,one is used to extract event elements,and the connection condition random field is used to get the best tag sequence;Another branch is used to obtain fault types,connect attention layer to capture important local information,and finally get classification results through softmax layer.Secondly,Aiming at the power grid emergency plan recommendation problem,this thesis converts the emergency plan recommendation problem into a computational similarity problem,and proposes a method combining a power grid domain-specific dictionary and a BERT model to complete the text similarity matching task.First,fine-tune the text similarity through the BERT pre-trained model,and then calculate the correlation of the faulty device through the proprietary dictionary in the power grid field.The text similarity and the device similarity are added to obtain the grid fault similarity.Finally,the most similar fault case treatment method is selected as the final recommended solution.The method in this thesis adds semantically related calculations to the traditional method,making the model more realistic.In this thesis,a comparative experiment is performed on the Chinese corpus of power grid failure cases.The experimental results show that the proposed method is better than the existing methods.