Research On State Evaluation Of The Thyristor Converter Valve Based On Data Drive

At present,the maintenance methods of power equipment in China are gradually changing from traditional maintenance methods such as regular maintenance to state maintenance.Evaluating the state of power equipment and accurately judging its operating state is the basis and core content of achieving state maintenance.As the core equipment of UHV DC transmission,the UHV converter valve plays a key role in rectification,inversion,etc.If it can be accurately evaluated in the operating state,abnormal conditions can be detected in time,the failure rate can be reduced,and the operation safety level of the power grid can be improved.However,due to imperfect evaluation technical standards,there is still no reliable state determination system.Aiming at the problem of evaluating the operating state of UHV DC transmission thyristor converter valves,this dissertation mainly conducts the following four aspects of work:(1)Operation state analysis of the thyristor converter valve.Firstly,the key components,basic functions and working principles of the thyristor converter valve of ABB’s technical route are deeply studied in combination with the site diagram and schematic diagram.Then,based on the 39 characteristic state quantities considered in the "Guidelines",a linear complementary fusion thyristor converter valve feature selection method is proposed,that is,three different evaluation criteria feature selection methods are used to calculate the comprehensive importance of fault features,and the optimal feature combination was screened out.Finally,the state division method in the "Guidelines" was used to classify the operating state of the thyristor converter valve into four grades: normal state,attention state,abnormal state and severe state,supplemented by different maintenance arrangements.(2)A sample equalization method for state evaluation of thyristor converter valves based on adaptive weighted oversampling.Firstly,the traditional processing methods such as undersampling and oversampling are introduced,and their principles and defects are discussed respectively.Then,based on the principle of Synthetic Minority Oversampling Technique(SMOTE),a new adaptive weighted oversampling algorithm is proposed.The algorithm gradually divides minority sub-clusters through hierarchical clustering,and the sub-cluster misclassification rate determines the final sampling scale,avoiding problems caused by traditional SMOTE,such as over-fitting of the model and overlapping of synthetic samples and the majority class samples,which could effectively improve the recognition rate of decision boundary samples.Through case analysis,it is verified that the proposed model is effective in realizing sample equalization and data enhancement in the operating state evaluation of thyristor converter valves,and it is better than traditional SMOTE,Borderline SMOTE and other oversampling algorithms.(3)A state evaluation method of thyristor converter valve based on Boosting algorithm.Firstly,the basic idea of Boosting algorithm in ensemble learning is explained,and two classification models XGBoost and LightGBM are introduced.Then,cross-validation and grid search are combined to find the optimal hyperparameters in each model.Afterward,based on the evaluation indicators such as accuracy rate,recall rate,and running time,the optimal Boosting algorithm is selected,and the final state evaluation model of the thyristor converter valve is constructed.Finally,on the basis of the equalized data generated in Chapter 3,the accuracy and effectiveness of the proposed thyristor converter valve state evaluation method are verified by an example.(4)An interpretability analysis method based on SHapley Additive ex Planations(SHAP)for the state evaluation model of thyristor converter valve.In order to clarify the specific operation logic of the evaluation model,the concept of machine learning model interpretability is introduced.The properties and application scenarios of Shapley value and SHAP attribution theory are studied,and three analysis methods of the converter valve state evaluation model are proposed,including global importance interpretation,feature distribution correlation interpretation,and individual decision interpretation.In the calculation example,the state evaluation model selected in Chapter 4 is analyzed,and the global and individual explanations are aided by drawing a summary diagram,a dependency diagram,and a force diagram.The concepts of faithfulness and correctness are introduced,and the interpretation models are compared to verify that SHAP attribution theory is more accurate and reliable than the traditional interpretation in converter valve state evaluation.