Physics-Data Dual-driven Online Calculation Method for Losses and Hotspot Temperature of High-power High-frequency Transformers
摘要
损耗与温升直接影响大容量高频变压器的绝缘寿命与运行可靠性,精确的热状态感知对其安全运行至关重要.为了实现高频变压器温度及损耗的快速精准在线监测,提出了一种基于物理-数据双驱动的在线计算模型.首先,构建考虑热传导频变效应与热各向异性的改进热网络模型,验证了其具有与有限元仿真相当的高精度;然后,针对实验样本难以测得,而有限元样本生成耗时长的问题,提出了物理约束生成对抗网络(physical constraint wasserstein generative adversarial network with gradient penalty,PC-WGAN-GP),生成了符合热力学定律的高质量数据集;最后,构建了基于冠豪猪优化算法(crown porcupine optimization,CPO)优化的物理信息卷积神经网络(physics-informed convolutional neural network,PI-CNN)模型,通过引入动态权重的混合损失函数,实现了物理机理约束与数据驱动特征的深度融合.研究结果表明,改进热网络模型对关键节点热点温度的计算相对误差在3.31%以内,确立了等效物理基准;在此基础上,所提出的 CPO-PI-CNN 模型在测试集上的相对平均绝对误差降低至2.66%,决定系数 R2达 99.19%,验证了该方法具备良好的计算精度与泛化能力.
Losses and temperature rise directly affect the insulation life and operational reliability of high-power high-frequency transformers.Accurate thermal state perception is crucial to its safe operation.In order to realize fast and accurate on-line monitoring of temperature and loss of high frequency transformer,an on-line calculation model based on physical-data dual drive is proposed.Firstly,an improved thermal network model considering the frequency-dependent effect of heat conduction and thermal anisotropy is constructed,and its high accuracy is verified to be comparable to that of finite element simulation.Then,aimed at the problem that the experimental samples are difficult to measure and the fi-nite element sample generation takes a long time,a physical constraint wasserstein generative adversarial network with gradient penalty(PC-WGAN-GP)is proposed to generate a high-quality data set that conforms to the law of thermody-namics.Finally,a physics-informed convolutional neural network(PI-CNN)model based on crown porcupine optimization(CPO)is constructed.By introducing a hybrid loss function of dynamic weights,the deep fusion of physical mechanism constraints and data-driven features is realized.The results show that the relative error of the hot spot temper-ature of the key nodes calculated by the improved thermal network model is within 3.31%,and the equivalent physical benchmark is established.On this basis,the relative mean absolute error(RMAE)of the proposed CPO-PI-CNN model on the test set is reduced to 2.66%,and the coefficient of determination R2 is 99.19%,which verifies that the method has good calculation accuracy and generalization ability.