A fault diagnosis method for converter transformers of UHVDC converter station using LightGBM-BTCN-MCAM with dissolved gas analysis

• A unified LightGBM-BTCN-MCAM framework is developed for intelligent fault diagnosis of UHVDC converter CTs, substantially improving the accuracy, robustness, and interpretability of diagnostic results. • Feature selection and redundancy reduction are achieved through LightGBM-based importance quantification, which enhances both diagnostic efficiency and model generalization. • A BTCN with skip connections is proposed, leveraging Bayesian and temporal convolutional layers and skip connections to extract hierarchical feature representations for robust fault diagnosis. • The MCAM is introduced to dynamically assign weights to key feature channels, thereby strengthening sensitivity to critical diagnostic cues and mitigating the influence of irrelevant noise. Converter transformers in ultra-high voltage direct current (UHVDC) converter stations are highly susceptible to faults due to prolonged exposure to extreme electrical, thermal, and environmental stresses. To address this challenge, this paper proposes a novel fault diagnosis framework within LightGBM-BTCN-MCAM, based on dissolved gas analysis (DGA). The framework first augments diagnostic features using gas concentration data from transformer oil. Subsequently, the light gradient boosting machine (LightGBM) is employed to quantify feature importance, enabling effective feature selection, mitigating redundancy, and improving computational efficiency. A Bayesian temporal convolutional network (BTCN) with skip connections is then designed to capture hierarchical temporal-spatial dependencies and extract robust fault representations. Furthermore, a multi-channel attention mechanism (MCAM) adaptively reweights feature channels, highlighting the most discriminative diagnostic information to enhance accuracy. Experimental evaluations demonstrate that the proposed LightGBM-BTCN-MCAM framework achieves superior diagnostic accuracy and robustness under noisy and imbalanced data conditions, while maintaining high efficiency. These results confirm its suitability for practical deployment and its potential to significantly enhance the reliability and safe operation of UHVDC converter transformers in complex real-world environments.