Prediction Of Dissolved Gas Concentration In Transformer Oil Based On Deep Learning

In order to achieve the strategic goal of"national interconnection"in my country's power sector,long-distance,UHV,and large-capacity power transmission modes have been widely used.As one of the core equipment in the power system,the power transformer plays an important role in voltage and current changes,and affects the stable operation of the power grid.Considering that there is a close relationship between the operating conditions of the transformer and the dissolved gas content in the internal oil,the accurate prediction of the dissolved gas in the oil can detect potential latent faults in the transformer equipment early,and provide theoretical support for the maintenance of the operators.This paper takes the transformer oil chromatographic data monitored online as the research object,combined with deep learning algorithms,and mainly conducts the following two researches:Aiming at the problem of"abnormal value"or"missing value"in the dissolved gas data extracted from the power transformer in normal operation,this paper proposes a transformer online based on a stacked noise reduction autoencoder(SDAE)Monitoring oil chromatography data cleaning model.Using online monitoring of the dissolved gas data in the oil of a 220k V and 500k V voltage grade transformer as the research object,first input the transformer oil chromatographic data collected under normal operating conditions into the SDAE cleaning model,and train the model to determine its key The size of the parameter;then input the oil chromatographic data that has"missing values"or"abnormal values"under normal operating conditions into the cleaning model,and compare the reconstruction error N_L with the loss function peak T_op,tolerance time window T_n,and error persistence The size relationship between the time T_t judges the type of bad data;finally,the data with"missing values"and"outliers"is reconstructed and repaired,and more original data information is restored,and the characteristic gas concentration is determined for the follow-up of the article.Accurate forecast research provides a reliable source of data.In view of the difficulty of determining key parameters and low prediction accuracy of traditional deep learning prediction models,this paper proposes a combination of dissolved gas concentration in oil based on improved particle swarm optimization(IPSO)and gated circulation unit(GRU)Forecast model.Divide the collected 7 characteristic gas sequences into training set and test set,use IPSO algorithm to iteratively search for the optimal values of key parameters of the GRU model(number of neurons m and learning rate?)on the training set data,and establish an IPSO-GRU combination the predictive model is used to verify the reliability of the proposed method on the test set data.The experimental results show that the method proposed in this paper not only overcomes the problem of selecting key parameters of the model,Performing feature data sequence cleaning improves the accuracy of the prediction model,but also has higher prediction accuracy than the traditional PSO-GRU,GRU,LSTM,RF and RNN deep learning algorithm models.The number of training sets continues to increase,and the prediction accuracy of the model gradually improves,indicating that the model can obtain more potential laws among the characteristic data sets,which plays an important role in timely and accurate judgment of the operating status of the power transformer.