【作者】 李峰；

【导师】 宋爱波；

【作者基本信息】 东南大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 现代社会生产力的发展对电网供电可靠性的要求越来越高,电网系统发生故障时如何准确地判断电网故障设备是电网系统故障诊断面临的重大挑战。然而传统电网故障诊断方法仍然存在以下缺陷:(1)没有充分利用电网结构信息,导致诊断方法可移植性较差;(2)输入特征依赖人工经验、抽象程度低,导致诊断方法的鲁棒性较差;(3)模型逻辑复杂,多分类的诊断模型等问题,导致难以解决电网多故障诊断问题。针对以上问题,本文研究基于深度学习的电网故障诊断技术,主要工作如下:针对传统电网故障诊断方法中存在可移植性和鲁棒性较差的问题,本文设计了一种自适应的电网高级特征提取方法。采用新颖的电网结构知识表示方式,将电网结构转化为特殊的图结构。在此基础上,高效地利用电网结构和故障告警信息,采用迭代更新的方式,设计了适用于电网结构的深度迭代网络(Deep Iterative Network),用于提取结构相关的抽象特征,增强了故障诊断方法的可移植性和鲁棒性。针对传统电网故障诊断方法难以有效解决电网多故障诊断问题,本文在深度迭代网络提取特征的基础上,设计了基于监督学习的电网故障诊断模型(AFD_SL)和基于强化学习的电网故障诊断模型(AFD_RL)。其中AFD_SL采用逐一诊断的方式设计模型,适用于有标签数据充足的电网故障诊断场景;AFD_RL将故障诊断问题转化为序贯决策过程,适用于缺乏有标签数据的电网故障诊断场景。综合使用上述两种模型,可在不同场景下高效地解决电网故障诊断问题。在实验室的高性能计算平台上,基于IEEE标准电网结构生成的仿真数据进行实验测试与分析,并与传统电网故障诊断方法进行对比。实验结果表明,本文设计的电网故障诊断方法在较好地处理多故障诊断问题的同时,具有良好的模型可移植性和鲁棒性。更多还原

【Abstract】 The development of modern social productivity demands better reliability of grid power supply.How to accurately determine the grid fault component when the grid system fails is a major challenge for grid system fault diagnosis.However,the traditional power grid fault diagnosis method still has the following defects:(1)The power structure information is not fully utilized,resulting in poor portability of the diagnostic method;(2)The input characteristics depend on artificial experience and with low degree of abstraction,which leads to poor robustness of the diagnostic method;(3)The complex logic of the model and the multiclassification diagnosis model make it difficult to solve the multi-fault diagnosis problem of power grid.In view of the above problems,this paper studies the power grid fault diagnosis technology based on deep learning.The main works are as follows:Aiming at the poor portability and poor robustness of traditional power grid fault diagnosis methods,this paper designs an adaptive feature extraction method of power grid.The grid structure is transformed into a special undirected graph structure by adopting a novel knowledge representation of the grid structure.On this basis,the grid structure and fault alarm information are used efficiently,and the iterative update method is adopted to design a Deep Iterative Network(DIN)which is suitable for grid structure,which is used to extract structural abstract features and enhance portability and robustness of fault diagnosis method.Traditional power grid fault diagnosis method is difficult to effectively solve the problem of power grid multi-fault diagnosis.Based on the feature extracted by deep iterative network,this paper designs a power grid fault diagnosis model based on supervised learning(AFD_SL)and a power grid fault diagnosis model based on reinforcement learning(AFD_SL).AFD_SL uses one-by-one diagnosis to design the model,which is suitable for power grid fault diagnosis scenarios with sufficient labeled data;AFD_RL transforms the fault diagnosis problem into a sequential decision-making process,which is suitable for power grid fault diagnosis scenarios without labeled data.The above two models can be used in different scenarios to solve the problem of power grid fault diagnosis efficiently.On the high-performance computing platform of the laboratory,with the experimental data generated based on the IEEE standard grid structure,the power grid fault diagnosis model designed by this paper is tested and analyzed,and compared with the traditional grid fault diagnosis method.The experimental results show that the power grid fault diagnosis methods designed in this paper have good model portability and robustness while dealing with multifault diagnosis problems.更多还原