Islanded microgrid formation (IMF) enhances active distribution network resilience but is challenging for current mathematical optimization methods due to its mixed-integer nonlinear nature. Deep reinforcement learning (DRL) also struggles with complex IMF problems due to action space explosion and environmental nonstationarity. This article proposes a multiagent DRL (MADRL) method for IMF, incorporating probability space optimization and random sequential updating to address IMF problems with numerous switches. Specifically, a multiagent framework models each controllable switch as an independent agent, and a specific actor-critic network architecture is designed for discrete control problems. To achieve unbiased policy gradient estimation, a probability space optimization loss function is devised to replace the Gumbel-Softmax-based gradient estimation in existing action-aware discrete DRL algorithms. Combined with the random sequential updating mechanism, the environmental nonstationarity issue faced by each agent is effectively mitigated. This results in a discrete MADRL-based IMF strategy with high computational efficiency and stable convergence, even when numerous switches are involved. Case studies on a modified IEEE 123-node system demonstrate that this method achieves optimality ratios of 96.85% and 97.41% in 16-switch and 23-switch scenarios, respectively.

Xinlin Zhuang, Jiahui Peng, Ren Ma, Yinfan Wang, Tianyi Bai, Xingjian Wei, Qiu Jiantao, Chi Zhang, Ying Qian, Conghui He. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Abstract Nowadays, the volume of literature in each field is huge and is growing rapidly, which posts challenge to researchers’ literature review. In this circumstance, developing useful tool for achieving efficient literature management is of high value. Traditional literature management tools, such as tools for key word searching, paper recommendation, relation visualization, and keyword cloud drawing, are not suitable for conducting content-level literature review. To address the issues of traditional literature management tools, a novel problem and method-oriented fine-grained academic graph is proposed to facilitate the exploration of research questions, methodologies, study perspectives, and their connections hidden in massive literature. For building such graph, a new ontology dedicated for describing the features of research paper is developed, an innovative multi-relation join extraction model is proposed, and a creative approach for leveraging the Large Language Models (LLM) to augment the triplet extraction results generated by supervised-learning model is developed. Experiments on widely used benchmark datasets show that the proposed multi-relation extraction model is able to achieve at least 8.01% and 8.65% improvement on entity identification and relation classification respectively, compared with state-of-the-art models. The visualized demonstration of the proposed graph shows that our graph is capable of accurately capturing the problem network, method network and hot topics hidden in massive literature. The Q&A system supported by the proposed graph demonstrates that our graph is really helpful for conducting literature review.

Power systems are well-engineered systems designed to supply power generation and services to end users. They are considered as the most crucial infrastructures in modern societies as numerous facilities such as telecommunication, transportation, public health, and emergency services heavily rely on a continuous power supply for their normal operation. The complex structure and interconnection with other facilities make power systems vulnerable to external threats such as natural disasters, extreme weather, and vandalism. On the other hand, with the advancements of artificial intelligence in the recent years, various cutting-edge control and optimization methods are emerged. Among them, deep reinforcement learning (DRL) is the most promoted for enhancing power system resilience. In this context, this article performs an in-depth review of various DRL methods and their applications in resilience optimization and enhancement of power systems. The review offers a structured examination of DRL methodology, along with a systematic categorization of existing literature into three themes (proactive actions, emergency response, and restoration and recovery) according to their contingency stages. The strengths and limitations of DRL-based resilience enhancement strategies are discussed across robustness, scalability, interpretability, and safety. A research roadmap is provided to highlight possible avenues for further exploration.