Collaborative Research: FMitF: Track I: Towards Verified Robustness and Safety in Power System-Informed Neural Networks

神经网络（NN）彻底改变了我们运营和管理现代电力系统的方式，为复杂的非线性关系建模提供了卓越的解决方案，并使用最先进的监控传感器收集的大量数据执行模式识别任务。尽管这些模型具有很好的优势，但其效率和可靠性可能会受到噪声或有偏见的功率测量以及可再生能源的不可预测性的负面影响。基于NN的模型由于其固有的非线性，高维性质和对抗性攻击的脆弱性而进一步复杂化。认识到与缺乏正式的鲁棒性保证的经验方法相关的风险，特别是在模型故障可能导致灾难性的现实后果的领域，该项目旨在通过优化针对现代电力系统特性定制的尖端NN验证器（alpha，beta-CROWN）来提高电力系统的安全性和可靠性。由此产生的改进旨在为电网运营商提供安全，可靠的工具来操作电力系统。此外，本项目还将为本科生和硕士生提供机器学习和电力系统领域的教育和研究支持。本项目旨在弥合电力系统与强大的神经网络验证技术之间的现有差距，分为三个重点。在Thrust I中，该项目将把NN验证器的应用扩展到拓扑感知的电力系统，检查不同的场景，包括对各种模型结构的完整和不完整验证，并相应地调整分支和边界算法。推力II将提高目前的神经网络验证器的有效性，结合电力系统的静态和动态约束，并进一步提高验证效率，通过认证培训。最后，在Thrust III中，该项目将为电力系统开发专门设计的验证器，作为基于灵敏度分析的电力系统规划的新工具。这最后一个组成部分首次采用了验证方法，利用电力系统内可解释的人工智能。总的来说，这些研究成果将彻底改变人们对电力系统正式鲁棒性验证技术的理解和应用，确保现代电力网络的安全性和可靠性。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Neural Networks (NNs) have revolutionized the way we operate and manage modern power systems, providing remarkable solutions to modeling complex non-linear relationships and performing pattern recognition tasks using abundant data collected by state-of-the-art monitoring sensors. Despite the promising advantages, the efficiency and reliability of these models can be negatively impacted by noisy or biased power measurements and the unpredictability of renewable energy sources. The NN-based models are further complicated by their inherent non-linear, high-dimensional nature and vulnerability to adversarial attacks. Recognizing the risks associated with empirical methods that lack formal robustness guarantees, especially in a field where model failures can lead to disastrous real-world consequences, this project seeks to enhance the security and reliability of power systems by optimizing the cutting-edge NN verifier (alpha, beta-CROWN) tailored to the characteristics of modern power systems. The resulting improvements aim to provide power grid operators with safe, dependable tools to operate the power systems. Moreover, this project also intends to support education and research initiatives, encompassing the fields of machine learning and power system, for both bachelor's and master's degree students.With a vision to bridge the existing gap between the power systems and the robust neural network verification techniques, this project is divided into three thrusts. In Thrust I, the project will extend the applications of NN verifiers to topology-aware power systems, examining different scenarios that include complete and incomplete verification on various model structures and adjusting branch and bound heuristics accordingly. Thrust II will enhance the effectiveness of current NN verifiers by incorporating power system static and dynamic constraints and further improve verification efficiency through certifiable training. Lastly, in Thrust III, the project will develop specially designed verifiers for power systems to serve as a novel tool for sensitivity analysis-based power system planning. This last component incorporates verification approaches for the first time, utilizing explainable Artificial Intelligence within power systems. Collectively, these research efforts will revolutionize people’s understanding and application of formal robustness verification techniques to power systems, ensuring the security and dependability of modern power networks.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.