Coordinated Power Grid Protection based on Machine Learning Methods

脱碳的过程导致电网中新型电网设备的大量增加，例如基于可再生能源、储能和高动态负荷的易失性馈电系统。此外，配电网的高分段或多分段布置、输电网临时高负荷的救治性再调度等非常规网络结构和电网运行方式也越来越多。协调电网保护以保持供电可靠性和电网安全的迫切需求仍然存在。由于运行和故障场景的分离越来越不清晰，以及网络结构和电网运行模式的多样性，经典的电网保护方法已经达到了极限。即使是适应性保护概念也不能充分应对这一挑战。在申请的研究项目中，将采用一种全新的电网保护方法。将基于神经网络的机器学习方法的非线性分类、能力学习和泛化能力等基本特性巧妙地运用到电网保护技术中，形成一种普遍适用的自动化电网保护解决方案。过流保护、距离保护或差动保护之间的功能区别在这里不再有效。以前的保护协调规划过程也正在成为一个神经网络结构的自动化训练过程，具有标记的真实和模拟运行和故障数据的时间序列。经过适当训练的特工随后在现场更换防护装置。集中或分散的解决方案都是可能的。利用“已知算子学习”方法将电网的物理知识引入神经网络结构，保证了保护技术对鲁棒性和可追溯性的重要要求。新方法将在具有数字开关设备配置的电网实验室模型上实施和测试，并借助实时网络模拟器（RTDS）。本课题的实施将为未来电网供电可靠性和网络安全的维护做出重要贡献，并促进电网的数字化进程。

The process of decarbonization leads to a massive increase in new grid devices in the electrical power grid, such as volatile feed-in systems based on regenerative energy sources, energy storage and highly dynamic loads. In addition, there are more and more unconventional network structures and grid operating modes, such as high degrees of meshing or multi-leg arrangements in the distribution grids and curative redispatch with temporary higher loadings in the transmission grids. The urgent need for coordinated grid protection to maintain supply reliability and grid security remains. Due to the increasing lack of clarity in the separation between operational and fault scenarios as well as multivariate network structures and grid operating modes, classic grid protection methods are reaching their limits. Even adaptive protection concepts cannot adequately meet this challenge in its full form. In the research project applied for, a fundamentally new grid protection approach will be taken. The basic properties of non-linear classification, competence learning and the generalization ability of machine learning methods based on neural networks should be used skillfully in grid protection technology and lead to a universally applicable and automated grid protection solution. The functional distinction between overcurrent protection, distance protection or differential protection is no longer effective here. The previous planning process of protection coordination is also becoming an automated training process for neural network structures with labeled time series of real and simulated operating and fault data. Appropriately trained agents then replace the protective devices in the field. Centralized or decentralized solutions are possible. The introduction of physical knowledge about the electrical grid into the neural network structure with the "Known Operator Learning" method ensures the important requirement of protection technology for robustness and traceability. The new approach is to be implemented and tested on a laboratory model of a power grid with digital switchgear configuration and with the help of a real-time network simulator (RTDS). With the implementation of this research project, an important contribution is made to maintaining the reliability of supply and network security in future networks and, last but not least, the digitization of the networks is promoted.