| 引用本文: | 翁宗龙,李滨,肖佳文,张佳乐,白晓清.基于数据-物理模型融合驱动的原始-对偶自监督学习最优潮流求解方法[J].电力自动化设备,2025,45(4):202-208 |
| WENG Zonglong,LI Bin,XIAO Jiawen,ZHANG Jiale,BAI Xiaoqing.Primal-dual self-supervised learning optimal power flow solution method driven by data-physical model fusion[J].Electric Power Automation Equipment,2025,45(4):202-208 |
|
| 摘要: |
| 随着新型电力系统的构建以及清洁低碳能源体系的转变，高维强非线性、高不确定性、强耦合等特点使得现有最优潮流计算的复杂度急剧增加。基于数据-物理模型融合驱动，提出一种内嵌交流潮流方程的原始-对偶自监督学习的最优潮流求解方法。建立原始神经网络和对偶神经网络，并采用类增广拉格朗日的方法进行联合训练。原始神经网络仅预测所有节点的电压，在该训练网络中内嵌交流潮流方程，以计算发电机的有功和无功出力；对偶神经网络预测拉格朗日乘子估计值。仿真结果表明，所提方法不仅关注大量数据的底层特征，还优化解的质量，有助于更好地探索数据的结构和特性。同时，该方法无须预处理标签样本数据集，其计算精度和可信度优于数据驱动方法，其计算速度比传统物理模型驱动方法快数十倍。 |
| 关键词: 数据-物理融合驱动 类增广拉格朗日 原始-对偶自监督学习 最优潮流 内嵌交流潮流方程 |
| DOI：10.16081/j.epae.202412028 |
| 分类号:TM73 |
| 基金项目:国家自然科学基金资助项目(51967001)；广西自然科学基金资助项目(2020GXNSFAA297117) |
|
| Primal-dual self-supervised learning optimal power flow solution method driven by data-physical model fusion |
|
WENG Zonglong, LI Bin, XIAO Jiawen, ZHANG Jiale, BAI Xiaoqing
|
|
Guangxi Key Laboratory of Power System Optimization and Energy Technology, Guangxi University, Nanning 530004, China
|
| Abstract: |
| Along with the construction of new power system and the transformation of clean and low-carbon energy system, the characteristics of high dimensionality, strong nonlinearity, high uncertainty, and strong coupling have led to a sharp increase in the complexity of existing optimal power flow calculation. Based on the fusion driven by data-physical models, a primal-dual self-supervised learning optimal power flow solution method with embedded AC power flow equations is proposed. A primal neural network and a dual neural network are built, and the class augmented Lagrangian method is adopted for joint training. The original neural network only predicts the voltage of all nodes, and the AC power flow equations are embedded in the training network to calculate the active and reactive output power of the generator. The dual neural networks predict the estimation value of Lagrange multiplier. The simulative results show that the proposed method not only focuses on the underlying features of a large amount of data, but also optimizes the quality of solutions, which helps to better explore the structure and characteristics of the data. Meanwhile, the method does not need to preprocess the labeled sample datasets, its calculation accuracy and credibility are superior to the data-driven methods, and its calculation speed is dozens of times faster than the traditional physical model-driven methods. |
| Key words: data-physical fusion driven class augmented Lagrangian primal-dual self-supervised learning optimal power flow embedded AC power flow equation |
