DGA results

Power transformers are critical components in a power system network and they are subjected to multifactorial stresses. The dissolved gas composition of an oil-filled transformer provides good indication of the transformer health. Numerous methods have been developed to interpret the gas data, both conventional and artificial intelligent based. There are issues with inaccurate diagnosis, high dependency on expert judgment and rigid diagnosis standards as transformer designs, materials and issues continue to evolve. This paper attempts to predict transformer fault using Multiclass Classification Models by Microsoft Azure ML Studio, based on Duval Triangle 1. Trained with 628 samples and validated with 269 samples, the One-vs-All Multiclass model achieved an overall accuracy of 90.3%. The model is then compared with 4 published works of similar nature. The output fault classes of this work are clustered accordingly to facilitate good comparison. Its performance in terms of prediction accuracy is superior when compared with SOM (73.6%) and SVM (78.5%) models, similar as Auto-WEKA Random Forest model (90.3%) and inferior when compared with ANFIS (97.7%). When validated with 122 datasets of inspected and diagnosed faulty transformers from IEC TC 10 database and published works, the overall accuracy is 76.2% for 7 output fault classes. The prediction accuracy advances to 91.8% when the fault classes are clustered into 4 fault types. Less output classes result in higher accuracy as there is lower probability for the diagnosis to fall into adjacent or neighboring classes. More training data is expected to improve the overall performance of the model.

电力变压器（Power transformers）是电力系统网络中的关键组件，其承受多因素应力作用。充油式变压器的溶解气体组分能够有效反映变压器的健康状态。目前已开发出多种气体数据解读方法，涵盖传统方法与基于人工智能的方法。随着变压器设计、材料与故障类型的不断演进，当前仍存在诊断精度不足、过度依赖专家判断以及诊断标准僵化等问题。本文基于杜瓦尔三角1法（Duval Triangle 1），借助微软Azure机器学习工作室（Microsoft Azure ML Studio）中的多分类模型开展变压器故障预测研究。该模型以628个样本进行训练、269个样本进行验证，其中一对多（One-vs-All）多分类模型的总体准确率达到90.3%。随后，本文将该模型与4项同类已发表研究进行对比。为便于开展有效对比，本文对本研究的输出故障类别进行了针对性聚类。与自组织映射模型（SOM，73.6%）、支持向量机（SVM，78.5%）相比，本模型的预测准确率更具优势；与Auto-WEKA随机森林（Auto-WEKA Random Forest，90.3%）模型的准确率持平，而劣于自适应神经模糊推理系统（ANFIS，97.7%）模型。利用来自国际电工委员会TC 10数据库（IEC TC 10 database）及已发表研究的122组经检测与诊断的故障变压器数据集进行验证时，该模型针对7个输出故障类别的总体准确率为76.2%。当将故障类别聚类为4种故障类型时，模型预测准确率提升至91.8%。输出故障类别越少，诊断结果落入相邻类别的概率越低，因此模型准确率越高。后续可通过扩充训练数据集进一步优化模型的整体性能。