基于模态分解和混合式CNN⁃GRUT的变压器油中溶解气体预测方法

Prediction method of dissolved gas in transformer oil based on modal decomposition and hybrid CNN⁃GRUT

为防止电力变压器出现运维不足或者过度运维的情况,对其运行状态进行评估和潜在性故障进行预测具有重要意义。DGA技术是对变压器状态进行评估的有效方法,而变压器的机械振动、油温等原因会导致油中溶解气体信号呈非线性趋势,非稳定特性;致使预测难度增加,甚至日常测量气体数据缺失导致以DGA技术为主的在线监测系统无法监测变压器状态。针对以上问题,本文应用EEMD分解气体浓度信号集,而EEMD产生的高频本征模态函数会增加预测难度和影响预测精度,使用WPD进一步将子信号模态函数分解,针对过去机器学习无法分离和解析浓度信号间时间关联性和蕴藏特性的难题,本文提出了混合式CNN⁃GRUT预测模型,分离气体浓度子信号当中的蕴藏特性,深度解析气体浓度子信号集当中的时间关联特性,迭代子信号重组得到油中溶解气体浓度信号预测值。实验结果得出,提出的CMD⁃CNN⁃GRUT预测模型相较于BP、Elman等混合预测模型,CMD⁃CNN⁃GRUT的预测平均绝对误差减少2244%和309%,并且结合实验证明了所提出的预测模型的有效性。

In order to prevent power transformers from under⁃or over⁃O&M,it is important to evaluate their operat⁃ing status and predict potential faults.DGA technology is an effective method to evaluate the state of transformers,and the mechanical vibration and oil temperature of transformers will cause the dissolved gas signal in oil to show a nonlinear trend and unstable characteristics.As a result,prediction difficulties increase,and even the lack of daily measurement gas data makes it impossible for online monitoring systems based on DGA technology to monitor trans⁃former status.In view of the above problems,this paper applies EEMD to decompose the gas concentration signal set,and the high⁃frequency eigenmode function generated by EEMD will increase the prediction difficulty and affect the prediction accuracy,and use WPD to further decompose the sub⁃signal modal function,aiming at the problem that machine learning cannot separate and analyze the temporal correlation and hidden characteristics between con⁃centration signals in the past,this paper proposes a hybrid CNN⁃GRUT prediction model to separate the hidden characteristics in the gas concentration sub⁃signal.The time correlation characteristics in the gas concentration sub⁃signal set were deeply analyzed,and the predicted value of the dissolved gas concentration signal in oil was ob⁃tained by iterative sub⁃signal recombination.The experimental results show that compared with BP,Elman and oth⁃er hybrid prediction models,the average absolute error of CMD⁃CNN⁃GRUT prediction is reduced by 2244%and 309%,and the availability of the suggested prediction model is proved by experiments.