基于深度神经网络的数据驱动潮流计算异常误差改进策略  

Improvement Strategy for Abnormal Error of Data-driven Power Flow Calculation Based on Deep Neural Network

作　　者：雷江龙 余娟[1] 向明旭 杨知方 杨燕[1] 李文沅[1] LEI Jianglong;YU Juan;XIANG Mingxu;YANG Zhifang;YANG Yan;LI Wenyuan(State Key Laboratory of Power Transmission Equipment&System Security and New Technology(Chongqing University),Chongqing 400044,China)

机构地区：[1]输配电装备及系统安全与新技术国家重点实验室(重庆大学),重庆市400044

出　　处：《电力系统自动化》2022年第1期76-84,共9页Automation of Electric Power Systems

基　　金：国家自然科学基金资助项目(52077016);重庆市自然科学基金资助项目(cstc2020jcyj-msxmX0315);中央高校基本科研业务费专项资金资助项目(2020CDJ-LHZZ-079)~。

摘　　要：在考虑不确定性的N-1安全校核、可靠性计算等需大规模重复潮流计算的场景中,基于深度神经网络(DNN)的数据驱动方法存在部分潮流变量误差异常的问题,影响潮流越限判别的准确率。对此,首先通过理论推导,分析DNN参数更新过程及数据标准化原理,发现该问题的重要成因之一为:DNN仅根据标准化误差迭代训练模型,未计及潮流变量的真实学习误差及工程实际的精度要求,无法及时针对误差异常的潮流变量调整DNN参数。然后,面向潮流计算提出基于动态学习权重的DNN自适应训练方法。该方法通过每轮迭代中验证集的真实学习误差、越限误判率及误差统计指标,确定各潮流变量的学习权重,有效降低数据驱动潮流计算的异常误差。最后,在IEEE标准算例和Polish 2383节点系统上仿真验证了所提方法的有效性。In the scenarios requiring large-scale repetitive power flow calculations, such as N-1 safety checks and reliability calculations considering uncertainty, the data-driven method based on deep neural network(DNN) has the abnormal error problem of some power flow variables, which affects the accuracy of the over-limit judgment of the power flow. To this end, through the theoretical derivation, the update process of DNN parameters and the principle of data standardization are analyzed. It is found that one of the important causes of this problem is that the DNN only trains the model iteratively based on standardized errors, and does not take into account the true learning error of the power flow variables and the actual accuracy requirements of the project. And it is impossible to adjust the DNN parameters in time for the power flow variables with abnormal errors. Then, a DNN adaptive training method based on dynamic learning weights is proposed for power flow calculation. This method determines the learning weight of each power flow variable through the true learning error, over-limit misjudgment rate and error statistical indicators of the verification set in each iteration, which effectively reduces the abnormal error of the data-driven power flow calculation. Finally,simulations on IEEE standard examples and Polish 2383-node system verify the effectiveness of the proposed method.

关 键 词：数据驱动 潮流计算 深度神经网络(DNN) 异常误差分析 学习权重 

分 类 号：TM744[电气工程—电力系统及自动化] TP183[自动化与计算机技术—控制理论与控制工程]