Dissolved gas in transformer oil forecasting for transformer fault evaluation based on HATT-RLSTM

Forecasting of dissolved gas concentration in transformer oil is important for judging the status trend of transformer in advance. The DGA sequence prediction results of the traditional deep learning methods have the time delay problem, making it difficult to provide accurate data for the next transformer diagnosis. To address this issue, a brand-new hybrid model based on Hierarchical Attention Network(HATT) and Recurrent Long Short-Term Memory Network(RLSTM) is proposed. Firstly, CNN is used to learn the correlation between extrinsic factors and target gas, fully considering the coupling relationship between the gases in oil based on three-ratio method in the Convnet phase. Secondly, RLSTM is used to completely train input vectors at the same time point in the space hierarchy. Thirdly, Hierarchical Attention Network is used to further mine the temporal relationship of different time points between RLSTM layers in the time hierarchy. To validate the effectiveness of the proposed HATT-RLSTM approach, prediction performance is evaluated. The RMSEs of HATT-RLSTM are less than 22.9% over the whole traditional deep learning methods, indicating that time delay phenomenon can be eliminated by the proposed new model.