基于经验小波变换和梯度提升径向基的变压器油中溶解气体预测方法  

Dissolved Gas Prediction In Transformer Oil Based on Empirical Wavelet Transform and Gradient Boosting Radial Basis

作　　者：张鹏 齐波 张若愚 邵梦雨 李成榕 ZHANG Peng;QI Bo;ZHANG Ruoyu;SHAO Mengyu;LI Chengrong(State Key Laboratory of Alternate Electrical Power System With Renewable Energy Sources(North China Electric Power University),Changping District,Beijing 102206,China;Qingdao Power Supply Company,State Grid Shangdong Electric Power Company,Qingdao 266002,Shangdong Province,China)

机构地区：[1]新能源电力系统国家重点实验室(华北电力大学),北京市昌平区102206 [2]国网山东电力公司青岛供电公司,山东省青岛市266002

出　　处：《电网技术》2021年第9期3745-3754,共10页Power System Technology

基　　金：国家自然科学基金联合基金项目(U1866603);中央高校基本科研业务费专项资金(2019QN124)。

摘　　要：电力变压器是电力系统的关键设备,其运行状态与电网稳定性密切相关。变压器油中溶解气体分析(dissolve gas analysis,DGA)是判断其运行状态的重要方法,预测变压器未来时刻的油中溶解气体含量,可以辅助运维人员判断变压器未来的运行趋势,提前掌握运行状态确保稳定运行。然而,由于油中溶解气体的产生机制复杂且受到变压器特殊运行工况、严苛运行环境、复杂电磁环境等因素的影响,油中溶解气体时间序列将呈现非线性和非平稳性的特征,传统的基于回归拟合模型的预测方法很难挖掘时间序列的这些特征,从而导致预测准确性较低,无法用于对变压器运行状态和故障的预测和诊断。为了解决上述问题,该文利用经验小波变换将具有非线性和非平稳的油中溶解气体时间序列分解为多个复杂度较低的分量,使得预测模型更易挖掘其变化特征,随后,以径向基函数神经网络作为基学习器构建了梯度提升径向基,将径向基函数神经网络的最佳逼近、避免局部最小等优点与梯度提升机强大的监督学习能力相结合,实现对油中溶解气体分解分量潜在规律的深度挖掘,并最终实现对油中溶解气体数据的精准预测。基于现场在运变压器对所提方法进行验证,结果表明:对于单台变压器预测准确率可达98.30%,对于某区域电网内的全体变压器准确率可提升9.01%,且可以实现对变压器故障的准确预测。Power transformer, the key equipment of the power system, is closely related to the stability of the power grid in its operation states. The dissolved Gas Analysis(DGA) is an important method to judge the operating condition of a transformer. The prediction of the dissolved gas content in the oil of the transformer can help the operation and maintenance personnel to judge its future operation trend, and grasp its operation condition in advance to ensure its stable operation. However, due to the complex generation mechanism of the dissolved gas in oil and the influences of the transformer special operating conditions, the harsh operating environment, the complex electromagnetic environment and other factors, the time series of the dissolved gas in oil shows nonlinear and non-stationary characteristics. The traditional prediction method based on the regression fitting model can hardly mine these characteristics of time series, as a result, low prediction accuracy may be there and unavailable to the prediction and diagnosis of the operation conditions and faults of the transformer. In order to solve the above problems, this paper uses the empirical wavelet transform to decompose the nonlinear and non-stationary time series of the dissolved gas in oil into multiple components with lower complexity, making it easier for the prediction model to mine its variation characteristics. Subsequently, the radial basis function neural network is used as the basis learning machine to construct the gradient lifting radial basis. The advantages of the radial basis function neural network, such as the best approximation and local minimum avoiding, are combined with the strong supervised learning ability of the gradient hoist to realize the deep mining of the potential law of the decomposition components of the dissolved gas in oil. Finally the accurate prediction of the dissolved gas data in oil is realized. The proposed method is verified based on the on-site transformer. The results show that the prediction accuracy of a single t

关 键 词：电力变压器 油中溶解气体分析 时间序列预测 经验小波变换 梯度提升机 径向基函数神经网络 

分 类 号：TM721[电气工程—电力系统及自动化]