Multiscale feature decoupling via VMD and dual-channel networks for dissolved gas prediction in transformers

Abstract

To address the impact of the non-linearity and non-stationarity of transformer oil dissolved gas concentration sequences on prediction accuracy, a multi-scale hybrid prediction framework based on VMD optimized by PSO and deep learning is proposed. By introducing a sliding average pooling operation method, the subsequences obtained from VMD decomposition are further decoupled into long-term trend and residual components, establishing a dual-channel heterogeneous modeling system. BiGRU is employed to capture the trend component, while TCN integrated with a multi-head self-attention mechanism is designed to enhance the dynamic representation capability of key time points in the residual component, enabling the collaborative optimization of multi-scale features. Experimental results show that the proposed model outperforms traditional methods in predicting dissolved gas $C_2{H}_2$ in transformer oil, with a reduction in RMSE by 0.2918, a decrease in MAPE by 23.78%, and an improvement in $R^2$ by 37.25%. Predictions of other gas components, such as $H_2$ and total hydrocarbons, further demonstrate that the model exhibits strong generalization performance across various gas components.