Collaborative Research: RI: Small: Deep Constrained Learning for Power Systems

在过去的二十年里，人工智能在计算机视觉和自然语言理解等多个学科取得了显着进展。该项目旨在利用强大的人工智能来改造人类建造的最大机器电网。事实上，在发电中大量可再生资源的整合提出了大量的计算挑战，特别是解决复杂的优化问题的频率增加。该项目提出了一种新的范式，深度约束学习，以解决这些大规模的优化问题，在真实的时间，同时确保高效和可靠的网格操作。如果成功，该项目可能从根本上改变电网的运营方式，并带来显著的经济和环境效益。深度约束学习（Deep Constrained Learning，DCL）是机器学习和优化的紧密结合，能够在真实的时间内为大规模非凸优化问题提供可靠的接近最优的解决方案。该项目有助于新的科学和工程知识沿着两个方向。它首先展示了DCL如何通过将优化的关键方法结合到深度神经网络的训练周期中，提供一种原则性的方法来适应深度学习中的硬约束。其次，它展示了如何利用领域知识进行模型简化，使DCL能够处理真实的电网的规模和复杂性。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。

In the last two decades, artificial intelligence has achieved remarkable progress in a variety of disciplines such as computer vision and natural language understanding. This project aims at leveraging robust artificial intelligence for transforming the electrical power grid, the largest machine built by humankind. Indeed, the integration of substantial renewable resources in power generation raises substantial computational challenges and, in particular, the solving of complex optimization problems with increased frequency. The project proposes a new paradigm, Deep Constrained Learning, to solve these large-scale optimization problems in real time, while ensuring efficient and reliable grid operations. If successful, the project may fundamentally transform how the grid is operated and bring significant economic and environmental benefits. While the development of Deep Constrained Learning is grounded in energy applications, the project findings may generalize to a broader class of engineering applications with hard physical or operational constraints.From a scientific standpoint, Deep Constrained Learning (DCL) is a tight integration of machine learning and optimization that delivers, in real time, reliable near-optimal solutions to large-scale nonconvex optimization problems. The project contributes to new scientific and engineering knowledge along two directions. It first demonstrates how DCL provides a principled way to accommodate hard constraints in deep learning by combining key methodologies from optimization into the training cycle of deep neural networks. Second, it shows how to exploit domain knowledge for model reduction, allowing DCL to handle the size and complexity of real power grids.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.