基于缺失数据填补的变压器油中溶解气体体积分数预测OA

The prediction of dissolved gas concentration in oil is of great significance for earlyfault diagnosis of transformer.To address the issues of high data integrity requirements and weak generalization ability in current pre-diction algorithms for dissolved gas concentration in oil,in this paper the oil immersed transformers is taken as the research object anda prediction algorithm for dissolved gas concentration in oil based on missing data imputation and an improved Kalman filtering is proposed.First,a full probability model for missing data imputation is con-structed based on Bayesian theory and Gaussian prior assumption.The missing data imputation problem is trans-formed into solving the problem maximum likelihood estimation of model parameters,and the expectation maximiza-tion(EM)algorithm is employed to solve the modle,ultimately achieving the imputation of missing data.Then,Kal-man filtering is used for gas concentration prediction.To address the issue of divergence in Kalman filtering,a BP neural network is introduced and a feedback mechanism is established so to dynamically correct the prediction re-sults and improve prediction accuracy.The experimental results based on actual data show under a data missing rate of 30%,thecorrelation coefficient of the imputation results achieved by the proposed dmethod reaches 96.3%,which is 11.4%and 5.7%higher than those of the autoregressiveintegrated moving average(ARIMA)and extremely randomized trees(ERT),respectively.Meanwhile,the proposed method has a higher prediction accuracy,and the root mean square error(eRMSE)and mean absolute error(eMAPE)are reduced by 28.5%and 33.6%respectively compared to long short term memory(LSTM)and long short term memory-convolutional neural network(LSTM-CNN)methods.