摘要
为了解决传统物理信息神经网络(PINN)因同时确定神经网络参数和损失函数中机理模型平衡方程参数而造成的参数不准确性问题,提出了一种微调残差物理信息神经网络(Fine tuning Res-PINN)。Fine tuning Res-PINN结构可以认为是两个神经网络以残差结构进行串联,并以微调方式进行分步训练。在第一个神经网络中,根据深度学习的原理,由深层神经网络建立完整的黑箱模型并以均方根误差(MSE)为损失函数,以实现从输入到输出的近似映射;在第二个神经网络中,根据残差结构和微调的思路,建立以MSE和机理模型方程为损失函数的浅层物理信息神经网络,进一步对机理模型的参数进行整定。基于微调神经网络的训练方式,先训练深层神经网络,并冻结其参数后,再训练浅层物理信息神经网络。两个算例被用来验证Fine tuning Res-PINN的有效性。仿真结果表明,所训练的参数精确地接近实际参数。
To solve the problem of inaccurate parameters caused by determining the parameters of the neural network and the parameters of equilibrium equation of mechnism model in loss function at the same time in the traditional PhysicsInformed Neural Network(PINN),a fine-tuning residual physics-informed neural network named Fine tuning Res-PINN was proposed.The structure of the fine-tuning Res-PINN can be considered as two neural networks connected in series in the form of a residual network and trained in fine-tuning.In the first part,a complete black box model was established by a deep layer NN for an approximate mapping from the input to output,according to the principle of deep learning and with Mean Squared Error(MSE)as the loss function.In the second part,based on the structure of residual network and the idea of finetuning neural network,a shallow neural network was established with both MSE and mechanism model equation as loss function,which can further strengthen the parameter tuning of the mechanism model.Based on the fine-tuning neural network training method,the parameters of deep neural network were trained and frozen,and then the parameters of shallow physics-informed neural network and the parameters of mechanism model were trained.Two examples were used to verify the effectiveness of the Fine tuning Res-PINN.Simulation results show that the trained parameters are exactly close to the actual parameters.
作者
王海涛
王新超
朱颖
王钱超
潘蕾
WANG Haitao;WANG Xinchao;ZHU Ying;WANG Qianchao;PAN Lei(Power Generation Engineering Company,Jiangsu Power Design Institute Company Limited,China Energy Engineering Group,Nanjing Jiangsu 211102,China;Key Laboratory of Energy Thermal Conversion and Control of Ministry of Education,(Southeast University),Nanjing Jiangsu 210096,China)
出处
《计算机应用》
CSCD
北大核心
2022年第S02期175-179,共5页
journal of Computer Applications
基金
国家自然科学基金资助项目(51576040)。
关键词
物理信息神经网络
深度学习
残差神经网络
微调神经网络
燃气轮机
Physics-Informed Neural Network(PINN)
deep learning
Residual Neural Network(RNN)
fine-tuning neural network
gas turbine
