The electric power transformer is a critical component in electrical distribution networks, and the diagnosis of faults in transformers is an important research area. Frequency Response Analysis (FRA) methods are widely used for analyzing winding faults in transformers, particularly in Chinese power stations. However, the current approach relies on manual expertise to interpret FRA curves, which can be both skill-intensive and lacks precision. This study presents a novel approach using a Multilayer perceptron model to directly model and analyze FRA data, simulating various winding fault types and degrees in 12-disc winding and 10-disc winding transformers with different connection configurations, resulting in three distinct datasets. Six different Multilayer perceptron architectures were developed, with optimal models achieving recognition accuracies of over 99.7% for diagnosing fault degrees and more than 90% for fault types. Hence, this paper has yielded a model architecture that exhibits commendable performance in diagnosing various fault types and their severities in different models of transformers when utilizing different FRA connection methods. Additionally, a specialized diagnostic system called FRA-DiagSys with two-stage model utilization was developed, achieving 100% accuracy in diagnosing fault types and degrees for a specific winding-10 power transformer, surpassing other diagnostic methods and strategies.
翻译:暂无翻译