| Power transformers are important equipment to ensure the stable operation of transmission and distribution networks.Power failures caused by transformer failures will cause huge economic losses and social impacts.The content of dissolved gas in transformer oil is an important basis for the condition evaluation of transformers.Transformer faults are mostly potential faults,and online monitoring has a certain lag in the acquisition content of dissolved gas in oil.Therefore,short-term prediction(within seven days)of the dissolved gas in transformer oil can make up for the lag of online monitoring and find the potential fault of the transformer in advance,monitoring the insulation status of the transformer,has great significance to the reasonable arrangement of the condition-based maintenance of the transformer.The current methods for predicting dissolved gas content in transformer oil have the following problems:1)There are a large number of missing values in the dissolved oil content data,which disrupts the continuity of the online monitoring data time series and changes its data characteristics and changes trends,bring difficulties to the prediction work.2)The time series of dissolved gas in transformer oil has nonlinear and nostationarity.For the original sequence,it is difficult to discover the change law of the content of dissolved gas in the oil.3)Short-term prediction requires high accuracy.A single method for predicting dissolved gas in oil is not enough to explain time series with strong nonlinearity,and the accuracy is difficult to meet the requirements.At the same time,there are overfitting problems.However,the currently used combination prediction model for combining sub-model results is too rough,and the choice of the base learner is limited,can not give full play to the advantages of multiple models,affecting the accuracy of prediction.In view of the above problems,this paper proposes a short-term prediction method for the dissolved gas content in transformer oil based gradient boosting machine.First,this paper proposes a Markov model based method for interpolating dissolved gas monitoring data in transformer oil.The dissolved gas data in the oil that changes with time is converted into a Markov chain transferred between different states,and the complementary value of the dissolved gas data in the oil is calculated by using the forward and reverse state transition matrix.The similarity similarity between the completed data and the measured value can reach 99.999%.Compared with the conformal interpolation method and the mean interpolation method,the similarity has been increased by 5.32%and 4.57%.respectively.For extreme points and mutation points,the similarity between the completed data and the measured value can reach 98.956%.Compared with the conformal interpolation method and the mean interpolation method,the similarity has increased by 14.77%and 103.42%,respectively.Then,according to the time-frequency characteristics of the time series of dissolved gas content in oil,decompose it by empirical wavelet transform(EWT)to weaken the mutual interference between different scale information,and multiple time series components are obtained as the basis for subsequent prediction.The sample entropy is used to evaluate the decomposition effect.The sample entropy of the most complex time series component is only 36%of the original time series.Finally,this paper combine multiple base learners in the form of a gradient boosting machine,and establishe separate prediction models for different time series components.For the inherent shortcomings of the commonly used GBDT algorithm,radial basis function neural network(RBFNN)is used as the base learner of gradient boosting machine to establish a short-term prediction model of dissolved gas content in transformer.Finally,the method proposed in the paper is verified through actual cases.The results show that compared with the EWT-GBDT model and the EWT-LSTM model,the average relative errors of the EWT-GBRBFNN model predictions decrease by 51.43%and 22.72%.Correlations increase by 16.73%,6.61%,the standard deviation decrease by 41.84%,6.71%,and the coefficient of determination increase by 27.26%,2.21%.The method proposed in this paper is optimal in all aspects,and can better provide a basis for subsequent transformer condition evaluation and maintenance work. |
