Forecasting Dissolved Gas Concentration in Transformer Oil Using the AdaSTDM

Abstract

Accurately forecasting dissolved gas concentration (DGC) in transformer oil is crucial for ensuring the safety and reliability of power transformers and facilitating early anomaly warning. Current methods for forecasting DGC demonstrate limited effectiveness in non-stationary characteristics with data-distribution shifts. To address this, this paper presents a novel adaptive segmented temporal distribution matching (AdaSTDM) model, consisting of the Toeplitz inverse covariance-based clustering (TICC) algorithm and time distribution matching (TDM) algorithm. To effectively adapt to the different state distribution of the DGC data, the TICC algorithm is used to segment the state domain of the DGC sequence, and the Jensen-Shannon (JS) divergence is used as an indicator to evaluate the segmentation results. The TDM module is designed to mitigate data-distribution mismatches by learning common knowledge among different gas states. Experimental results across two real-world cases illustrate that the proposed AdaSTDM outperforms various advanced methods in predicting both stationary and non-stationary DGC data. © 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.