Scholarlink

<p>Oil-immersed transformers are essential components of power systems; however, they face significant risks from complex faults such as discharge-thermal (DT) conditions, which threaten electrical systems. Dissolved Gas Analysis (DGA) effectively detects faults by analyzing gas anomalies in insulating oil; however, traditional methods (e.g., Rogers, Duval Triangle) falter with intricate patterns, and conventional artificial intelligence models suffer from overfitting and bias owing to imbalanced datasets (e.g., rare faults occurring in less than 0.3% of samples). This study introduces a hybrid ensemble approach combining the synthetic minority over-sampling technique (SMOTE), optimized using five nearest neighbors to balance extreme imbalances (e.g., DT with eight DGA samples), and a gradient boosted decision tree for fault classification. Outperforming ADASYN and Borderline-SMOTE, SMOTE expands the imbalanced 2974 DGA sample dataset to 5152 balanced samples. Evaluated on 504 validation samples from 5152 balanced samples and tested on an independent 409 samples testing dataset, the method achieves 98.56% training accuracy and 98.53% testing accuracy. This approach exhibits outstanding performance in diagnosing complex faults, reducing misdiagnoses, and enhancing transformer reliability, thereby offering a practical tool for enhancing power system maintenance and reliability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Hybrid Ensemble Approach for Dissolved Gas Analysis-Based Power Transformer Fault Diagnosis Using SMOTE and GBDT

  • 期刊论文

  • 发表日期：

    2025年11月3日

查看全文
  • Huy Vu Tran,
  • Kim-Anh Nguyen

摘要

Oil-immersed transformers are essential components of power systems; however, they face significant risks from complex faults such as discharge-thermal (DT) conditions, which threaten electrical systems. Dissolved Gas Analysis (DGA) effectively detects faults by analyzing gas anomalies in insulating oil; however, traditional methods (e.g., Rogers, Duval Triangle) falter with intricate patterns, and conventional artificial intelligence models suffer from overfitting and bias owing to imbalanced datasets (e.g., rare faults occurring in less than 0.3% of samples). This study introduces a hybrid ensemble approach combining the synthetic minority over-sampling technique (SMOTE), optimized using five nearest neighbors to balance extreme imbalances (e.g., DT with eight DGA samples), and a gradient boosted decision tree for fault classification. Outperforming ADASYN and Borderline-SMOTE, SMOTE expands the imbalanced 2974 DGA sample dataset to 5152 balanced samples. Evaluated on 504 validation samples from 5152 balanced samples and tested on an independent 409 samples testing dataset, the method achieves 98.56% training accuracy and 98.53% testing accuracy. This approach exhibits outstanding performance in diagnosing complex faults, reducing misdiagnoses, and enhancing transformer reliability, thereby offering a practical tool for enhancing power system maintenance and reliability.