Probabilistic Modeling of Dissolved Gas Concentration for Predicting Operating Status of Oil-Immersed Transformers
Jianwei Shao, Junhua Wang, Xiao Pan, Ruilin Wang, Shuxun Liu, Zhanxiang Jin, Ziqing Wang
Abstract
Accurate prediction of dissolved gas concentration is vital for status assessment and fault diagnosis in oil-immersed transformers. Most current methods for predicting dissolved gas concentration produce deterministic results without providing any information on prediction uncertainty. This study presents for the first time a prediction model based on probability density, which integrates with interval estimation to assess the operating status of transformers. First, the quantile regression-long short term memory networks-Kernel density estimation model is developed to predict dissolved gas concentration and calculate the corresponding probability density function. The assessment of transformer status is then conducted using interval estimation, which includes a decision base consisting of prediction interval, attention value, and interval width. Experimental results illustrate that the proposed model achieves accurate predictions of gas concentration and probability density with excellent stability and applicability under various operating conditions, achieving the highest accuracy. Additionally, the findings reveal that the false alarm rate of the suggested criterion driven by uncertain information is 5.3% during normal operations and 0% under defective conditions. This article makes a contribution to the field by investigating uncertainty prediction models as valuable tools for predicting dissolved gas concentration and evaluating the condition of transformers.