基于轻量化混合门结构神经网络的变压器电磁场仿真方法

Transformer Electromagnetic Field Simulation Based on Lightweight Hybrid Gate Recurrent Unit Network

针对传统电力变压器电磁场仿真模型需要占用大量磁盘空间,运行功耗大,运算极为消耗计算资源等问题,提出一种基于轻量化混合门结构网络(lightweight hybrid gate recurrent unit network,LH-GRUNet)的变压器电磁场仿真方法。首先针对单相E芯变压器设计了一种混合门结构网络模型,实现了变压器铁芯和绕组的电磁场仿真;进一步,提出一种神经网络轻量化方法,通过删除模型全连接层中的冗余参数以压缩模型所需的存储空间与计算资源,最终实现变压器电磁场的低功耗、快速仿真。此外,基于LH-GRUNet的变压器电磁场仿真模型能够有效降低其运行所需的存储空间与计算资源,使其能够在几乎不损失模型精度的前提下应用于体积小、功耗低的边缘计算平台。

A Lightweight Hybrid Gate Recurrent Unit Network(LH-GRUNet)-based simulation for the transformer electromagnetic fields is proposed to address the problems of the traditional electromagnetic simulation models that require a large disk space,consume high power during the operation,and excessively use the computational resources.Firstly,a hybrid gate recurrent unit network model is designed for a single-phase E-core transformer to simulate the electromagnetic fields of the transformer's iron core and winding.Furthermore,a neural network lightweight method is proposed to compress the storage space and the computational resources required for the model by removing the redundant parameters from the model's fully connected layer,achieving the low power consumption and the fast simulation of the transformer electromagnetic fields.In addition,the transformer electromagnetic field simulation model based on the LH-GRUNet effectively reduces the storage space and the computational resources required for operation,enabling it to be applied to the small-volume and low-power edge computing platforms with almost no loss in the model accuracy.