基于注意力与长短期记忆网络的变压器代理模型  

Transformer surrogate model based on attention and long-short term memory

作　　者：金亮[1,2] 冯裕霖 曹佳豪 王艳阳[1] JIN Liang;FENG Yulin;CAO Jiahao;WANG Yanyang(Tianjin Key Laboratory of Advanced Technology of Electrical Engineering and Energy,Tiangong University,Tianjin 300387;State Key Laboratory of Reliability and Intelligence of Electrical Equipment,Hebei University of Technology,Tianjin 300130)

机构地区：[1]天津市电工电能新技术重点实验室(天津工业大学),天津300387 [2]省部共建电工装备可靠性与智能化国家重点实验室(河北工业大学),天津300130

出　　处：《电气技术》2021年第7期65-71,77,共8页Electrical Engineering

基　　金：国家自然科学基金面上项目(51977148)。

摘　　要：由于需要考虑换能效率、噪声、体积和质量等因素,电力变压器的设计参数和性能数据往往十分复杂,因此,如何建立变压器代理模型是亟需解决的问题。采用代理模型的优化算法(SBO)能有效解决数值模拟直接优化耗时长的问题。本文用深度学习建立变压器设计参数和性能数据的代理模型,实现变压器性能优化目标的高精度预测,有效降低变压器性能分析与优化所需时间。首先基于长短期记忆网络(LSTM)的深度学习模型,建立非晶合金变压器各个参数间的非线性映射,并加入注意力机制来增强模型的预测效果。最后,通过有限元仿真实验对提出的深度学习代理模型进行验证,并与其他常用的代理模型进行比较,证明了注意力机制与长短期记忆网络代理模型在预测精度方面的优越性。The design parameters and performance data of power transformers are often very complex considering many factors such as energy exchange efficiency,noise,volume and weight.Therefore,how to establish transformer surrogate model is an urgent problem to be solved.The surrogate-based optimization(SBO)can effectively solve the problem of long optimization time.In this paper,the surrogate model of transformer design parameters and performance data is established by deep learning to achieve high precision prediction of transformer performance optimization objectives and effectively reduce the time required for transformer performance analysis and optimization.Firstly,based on the deep learning model of long-short term memory network(LSTM),the nonlinear mapping between various parameters of amorphous alloy transformer is established,and the attention mechanism is added to enhance the prediction effect of the model.Finally,the proposed deep learning surrogate model is verified by finite element simulation experiment and compared with other commonly used surrogate models.The results show that the attention and long-short term memory surrogate model is superior in prediction accuracy.

关 键 词：非晶合金变压器 有限元方法 代理模型 深度学习 长短期记忆网络 注意力 

分 类 号：TM41[电气工程—电器]