Collaborative Research: Physics Informed Real-time Optimal Power Flow

这个NSF项目旨在利用机器学习技术开发一个物理信息实时最优潮流模型，以解决在为发电厂输出提供接近最优解决方案方面的差距，同时考虑实际的动态约束，以避免频率波动和电网不稳定。该项目的智力优势包括开发技术，将物理和动态原理集成到机器学习管道和方法中，以确保可扩展和可靠的解决方案来解决最优潮流问题。该项目的更广泛影响包括对电网的重大长期影响，减少碳排放和提高电网可靠性，特别是在极端天气，需求增加和间歇性发电的不确定性下。pi还将与国家实验室和非营利组织合作，以确保开发的模型可被更广泛的社区访问和使用，包括公用事业、政策制定者和研究人员。此外，该项目将为物理、工程和机器学习交叉领域的培训和教育提供机会，从而为能源和可持续发展领域的熟练劳动力的发展做出贡献。该项目做出了四个关键的科学和工程贡献：(1)将物理信息神经网络与传统前馈神经网络结合起来，实时预测最优潮流问题的解决方案，追求动态稳定性和最优性。(2)学习嵌入中保证约束满足的新方法。(3)研究确保训练可扩展性的技术，以适应大型、实际规模的网络。(4)通过评估模型在测量噪声下的性能和分析模型可靠性来追求模型的鲁棒性，从而深入了解最优潮流问题的高质量近似。提出的模型有望加快可再生能源进入电网的速度，减少因稳定性问题而导致的弃电和传统启发式下垂控制导致的次优。该奖项反映了美国国家科学基金会的法定使命，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。

This NSF project aims to develop a physics-informed real-time optimal power flow model using machine learning techniques to address the gap in providing close to optimal solutions for power plant outputs while considering practical dynamical constraints to avoid frequency fluctuations and grid instabilities. The intellectual merits of the project include developing techniques to integrate physical and dynamical principles in machine learning pipelines and methods to ensure scalable and reliable solutions to optimal power flow problems. The broader impacts of the project include significant long-term impacts on power grids, reducing carbon emissions and increasing grid reliability, especially under extreme weather, increased demand, and uncertainty from intermittent generation. The PIs will also engage with national laboratories and non-profit organizations to ensure that the developed model is accessible and usable by the broader community, including utilities, policymakers, and researchers. Furthermore, the project will provide opportunities for training and education in the intersection of physics, engineering, and machine learning, thereby contributing to the development of a skilled workforce in the field of energy and sustainability.The project makes four key scientific and engineering contributions: (1) Advancements in combining physics-informed neural networks with conventional feed-forward neural networks to predict solutions to optimal power flow problems in real-time, pursuing dynamic stability while also optimality. (2) Novel approaches of ensuring constraint satisfaction in the learned embedding. (3) Investigation of techniques to ensure scalability of training to large, realistically-sized networks. (4) Pursuit of model robustness by assessing model performance under measurement noise and analyzing model reliability to develop insights into high-quality approximations of the optimal power flow problem. The proposed model holds the promise to expedite the adoption of increased renewable energy into the power grid, reducing curtailment resulting from stability concerns and suboptimalities resulting from conventional heuristic droop control.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.