Research On Transformer Running State Prediction Based On Machine Learning

Power transformers,as one of the key equipment in the power system,undertake an important job.At present,with the development and construction of smart grids,state data and environmental data of power systems are gradually being integrated and shared on a unified platform.Traditional models based on theoretical analysis are difficult to deal with multi-dimensional,massive dataset information.In this context,starting from the inherent law of the data itself,using the data analysis method of machine learning,it provides a new solution and technical means for the state prediction of power transformers.Aiming at the problem of data noise and inconsistency,data preprocessing is required before transformer state prediction.In this paper,the correlation between parameters is considered,and the method of transformer state monitoring data cleaning is proposed based on sequence correlation analysis.First,a model for measuring the degree of association between state data is established to find a sequence with relevance.Then use the density-based clustering method to find out the missing data and abnormal data in the sequence.The abnormal data is classified into sensor abnormal data that can be cleaned and device state abnormal data that cannot be cleaned according to the data cleaning rule which considers the correlation.The data that can be cleaned is corrected by the wavelet neural network,and the time point is marked for the data that cannot be cleaned.The method proposed in this paper can identify abnormal data and missing data,and provide data support for subsequent transformer running state prediction.Aiming at the problem that the characteristic parameter trend prediction of transformer has poor stability and accuracy,this paper proposes a combined model for the short-term prediction of transformer state parameters.The kernel principal component analysis(KPCA)method is used to extract the main feature parameters from the multi-dimensional input vector as the input vector of the generalized regression neural network(GRNN)which reduces the influence of the irrelevant features on the model training.When training the GRNN,the improved fruit fly optimization algorithm is used to select the core parameters of the model.The example shows that the proposed method is simple in structure and can predict the change of short-term state parameters more accurately.Aiming at the problem that the transformer state parameter prediction has low accuracy in medium and long-term scale prediction,a deep long short-term memory network(LSTM)model is proposed.The model can effectively utilize historical data information due to the existence of the gating unit.The analysis of the example shows that the gas concentration prediction based on the deep LSTM network model improves the accuracy of the prediction,and also has high accuracy in the long-term concentration prediction problem,which can effectively realize the medium and long-term parameter trends prediction.Aiming at the problem of running state prediction of power transformers,a prediction method based on deep learning network is proposed by using machine learning method.Because the oil chromatographic data is easy to obtain and the sensitivity is high,the concentration of dissolved gas in the oil is selected as the input information of the network to predict the future trend of oil chromatography.Then the non-code feature values are calculated to provide more feature information for the model.The calculation results are used to train the deep belief network(DBN)to realize the classification of the running state.The experimental results show that the proposed method has higher accuracy and can accurately predict the changing trend of the transformer running state.The accurate prediction of the running state of the transformer can provide a scientific basis for rationally arranging the state maintenance of the transformer and ensuring the safe operation of the transformer.