近年来极端自然灾害频发，给电网造成巨大冲击。配电网与电力用户连接最为密切，提升配电网应对自然灾害的能力意义重大。本文围绕极端灾害下的配电网韧性提升优化决策方法展开研究，从灾害建模、事故仿真、灾前资源优化调配、灾后孤岛微电网供电及配电网抢修及重构等方面，综合提升配电网韧性。首先，从灾前角度，研究极端灾害和配电网灾后恢复特性对灾前人员物资购置调配的影响。考虑资源提供商、电网管理者、极端灾害之间的三方博弈，建立了灾前灾后协同式优化的扩展式博弈模型，考虑灾前资源购置、资源分配、抢修队伍配备及灾后配电网重构及抢修等多个阶段，利用逆推法逐层寻找非劣解进行模型求解。实验结果表明：灾后抢修的近似非劣解能较好逼近利用粒子群算法搜索得到的抢修方案，证明了抢修近似非劣解的有效性；在多故障场景下的测试，可以求得考虑灾后恢复特性的灾前资源购置方案，有效减小灾害导致的配电网失电损失。接着，从灾后自下至上式恢复角度，在配电网断电形成孤岛微电网期间，利用有限应急供电资源为其供电。研究了孤岛微电网的能量管理措施，最大化断电期间的供电效益，提升配电网灾后韧性。考虑了来自负荷需求、可再生能源出力、系统储能含量及断电时间的不确定性，基于OpenAI Gym和OpenDSS搭建了孤岛微电网的运行仿真平台。提出了双智能体强化学习的模型和训练算法，进行灾后孤岛微电网的电源出力控制和负荷动态切除。结果表明，本文方法在多个灾害场景、不同储能含量及配电网断电时间下均取得了较好的韧性提升效果。然后，从灾后自上至下式恢复角度，利用抢修及重构进行配电网灾后恢复。为应对灾情的不确定性、多抢修队伍调配、抢修与配网重构耦合等挑战，提出了适用于多抢修队调配的强化学习模型，设计了状态粗筛机制以固定强化学习状态及动作空间，得到适用于多故障场景的通用抢修调配策略。提出四种贪婪抢修策略作为对比，实验结果表明在提出的多个评价指标下强化学习综合表现稳定，可以作为灾后故障抢修的可选方案。最后，设计了韧性配电网事故演变及快速恢复仿真系统，集成了极端灾害建模、配电网事故仿真、灾前优化、灾中监控、灾后恢复等多项功能，进行各模块的输入输出数据格式设计和系统数据流设计。在CloudPSS上搭建了测试页面，接入了标准电网算例和真实电网算例，进行了各个功能模块的测试，工程实用价值较强。
In recent years, extreme natural disasters have occurred frequently, causing a huge impact on the power grid. Since the distribution system is closely connected with users, it is of great significance to improve the ability of the distribution system to respond to exemetre disasters. Focusing on the methods for improving the resilience of the distribution system under exemetre disasters, this article considers the aspects of extreme disaster modeling, accident evolution simulation, pre-disaster resource purchase and allocation, post-disaster island microgrid power supply, and distribution system emergency repair and reconfiguration to improve the resilience of the distribution system comprehensively.First, the method of purchasing and deployment of resources of the distribution system is studied to improve the resilience of the distribution system from the perspective of pre-disaster defense. Considering the three-party game relationship among resource providers, utilities, and extreme disasters, this article establishes a multi-stage extended game which includes pre-disaster resource purchase and allocation, emergency repair team deployment, disaster attack on the distribution system, and post-disaster distribution system reconfiguration and emergency repair to obtain the pre-and post-disaster collaborative optimization scheme of the distribution system. The efficient deriving method of non-inferior solution is designed layer by layer, and the inverse method is used to solve the game model efficiently. The experimental results under multiple failure scenarios show the proposed method can obtain the best pre-disaster resources purchase and deployment plan that takes into account the characteristics of post-disaster recovery, effectively reducing the expected power loss of the distribution system.Next, the microgrid is used to provide emergency power supply for the service interrupted distribution system in this article, the post-disaster energy management strategy of the service interrupted distribution system is studied to maximize the utility value during the power outage, thus improving the resilience of the distribution system. A dual-agent reinforcement learning model and algorithm are proposed for power output management on the source side and dynamic load shedding on the load side of islanded microgrids. Various uncertainties within the microgrid are considered, and an islanded microgrid operation simulation platform based on OpenAI Gym and OpenDSS is established. The results show that the proposed method can achieve a great resilience improvement effect in multiple disaster scenarios, different remaining energy storage, and power outage periods.Then, the post-disaster dynamic repair and reconfiguration plan of the faulty distribution system is studied to minimize the expected power loss, thus improving the resilience of the distribution system. In order to cope with the challenges from the uncertainty of the disaster, the deployment of multiple emergency repair teams, the coupling of emergency repair and reconfiguration, this paper builds a model for emergency repair deployment of multiple repair teams based on reinforcement learning. A state coarse screening mechanism is designed to fix the size of the state and action space of reinforcement learning. An environment based on OpenAI Gym is constructed for the repair and reconfiguration of the distribution system. The experimental results under different recovery modes and multiple failure scenarios show that reinforcement learning has a stable comprehensive performance under the proposed multiple evaluation indicators, and it can be used as an optional solution for dynamic repair after a disaster.Finally, based on the CloudPSS digital twin design platform, a resilient distribution system accident evolution and rapid recovery simulation system is designed, which includes multiple functions such as extreme disaster modeling, distribution system accident simulation, pre-disaster optimization, real-time monitoring during disasters, and post-disaster recovery. A real power grid case with tens of thousands of nodes is integrated into the system. The functions of the system are tested on real power grids and standard power grids. The results show that the system can effectively improve the resilience of the distribution system and has strong engineering practical value.
