基于改进物理信息神经网络的变电站建筑本体运行碳排放预测方法  

Carbon Emission Prediction Method for Substation BuildingsBased on Improved Physics Informed Neural Network

作　　者：高君玺 刘刚[1] 陈庆伟 王志鹏 GAO Junxi;LIU Gang;CHEN Qingwei;WANG Zhipeng(Tianjin University,Tianjin 300072,China;Economic and Technology Research Institute of State Grid Shandong Electric Power Company,Jinan 250021,China)

机构地区：[1]天津大学,天津300072 [2]国网山东省电力公司经济技术研究院,济南250021

出　　处：《建筑节能(中英文)》2025年第4期113-119,共7页Building Energy Efficiency

基　　金：国家电网有限公司总部管理科技项目资助(5200-202216099A-1-1-ZN)。

摘　　要：变电站建筑碳排放量的计算是一个复杂且具有挑战性的任务,尤其是在设计阶段预测碳排放量将有助于实现碳排放量的控制。神经网络(NN)模型由于其计算速度快、预测准确性高的优点被广泛应用。然而,实际应用中往往难以收集到模型所需的高质量数据集,这对模型的训练难度与预测鲁棒性造成了影响。事实上,可以通过总结数据集的先验信息,并通过先验信息与神经网络相结合提高模型的训练速度与数据集范围外的预测效果。提出一种用于碳排放预测的物理信息神经网络(PINN)构建方法,将遗传规划(GP)算法嵌入PINN结构中,以获取数据集的先验物理信息,从而提高模型性能。以山东省作为分析案例区域,对比两个模型在不同数据集下的训练结果,证明了在小数据集下相较于NN模型PINN具有更快的训练速度与更好的预测稳定性。Calculation of carbon emissions from substation is a complex and challenging task,especially in the design phase,which will help to realize the control of carbon emissions.Neural Network(NN)hmodels are widely used due to the advantages of fast computation and high prediction accuracy.However,it is often difficult to collect high-quality datasets required by the models in practical applications,which impacts the model training and prediction robustness.In fact,prior information of the dataset can be summarized and combined with neural networks to improve the training speed of the model and the prediction effect outside the range of the dataset.A Physics Informed Neural Network(PINN)hstructure is proposed for carbon emission prediction that embeds a Genetic Programming(GP)algorithm into the PINN to obtain prior physical information of the dataset to improve the model performance.Comparing the training results of the two models under different datasets from Shandong Province,the approach PINN is demonstrated to have faster training speed and better prediction stability than the NN model,especially in small datasets.

关 键 词：碳排放 物理信息约束神经网络 变电站 

分 类 号：TU271[建筑科学—建筑设计及理论] TM63[电气工程—电力系统及自动化]