Transformer Dissolved Gas Concentration Prediction Based on Quadratic Decomposition Reconstruction and BKA-BiLSTM

Abstract

For the prediction of each gas concentration in oil-immersed transformers, in this article, improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is first applied to the original gas data, and the sample entropy (SE) value of each subsequence is computed, the high-frequency sequences with the highest SE are subjected to quadratic variational mode decomposition (VMD) to further reduce the degree of its instability, that is, the ICEEMDAN-SE-VMD decomposition model is formed. Second, reconstruction operations are performed on the subsequences with close SE values after ICEEMDAN decomposition to reduce the prediction time while ensuring the accuracy. Finally, a bidirectional long short-term memory with attention mechanism (Bi-LSTM-AT) is used to predict each subsequence separately; for the optimization of the parameters in prediction algorithms, the latest black-winged kite algorithm (BKA) is used in this article for optimization of the parameters, and the prediction results of the subsequence are superimposed to be the final prediction value for the gas concentration. The prediction results of the six gases produced by the transformer show that compared with other methods, the prediction method used in this article reduces the prediction time, while the prediction accuracy is also guaranteed.