Prediction for Dissolved Gas in Power Transformer Oil Based on Transfer Learning and Mutation Detection

To solve the problem of poor prediction of dissolved gas concentration in transformer oil under defective state caused by "small-sample" problem, a prediction method for dissolved gas in power transformer oil based on transfer learning and mutation detection is proposed. Firstly, the domain-adversarial transfer learning method is used to train the prediction models under normal and defective states respectively. A large number of normal state samples are used to assist in training to solve the problem of inaccurate prediction caused by the "small-sample" characteristics of the defective state. Then, taking the prediction error as the criterion, a transformer state mutation detection method is proposed to determine the transformer state for selecting the required model during online prediction. Finally, the online monitoring data of dissolved gas concentration in transformer oil collected on site is used for testing. The results showed that the proposed method can adapt to the prediction tasks under normal and defective states at the same time. Compared with the comparison method, MAPE is reduced by more than 6.504%, and RMSE is reduced by more than 73.295 μL/L.