Prediction for Dissolved Gas in Transformer Oil Based on CVMD and TCN-GRU-KAN

Accurate prediction of dissolved gases in transformer oil is essential for evaluating the health of power transformers and preventing potential failures. However, monitoring data often contains strong noise, and traditional deep learning models exhibit limited feature extraction capability, leading to reduced prediction accuracy. To address these challenges, this paper proposes a novel hybrid forecasting model based on signal decomposition and integration. First, the Correlated Variational Mode Decomposition (CVMD) is applied to denoise the original gas time series and decompose it into intrinsic mode functions (IMFs) with distinct frequency characteristics. Next, an ensemble model integrating Temporal Convolutional Network (TCN), Gated Recurrent Unit (GRU), and Kolmogorov–Arnold Network (KAN) is used to independently predict each IMF component. The final prediction is obtained by aggregating the outputs of all subseries. Case studies demonstrate that the proposed method outperforms other algorithms, with the coefficient of determination (R2) increasing by 28.3%, indicating a significant improvement in prediction accuracy and generalization capability. These findings provide valuable technical support for online monitoring and early fault warning of power transformers.