 |
Center for Nonlinear Studies |
Renewable and distributed energy resources increase uncertainty and system complexity, challenging secure power system operation. Machine learning approaches including neural networks have shown substantial promise e.g. for power system security assessment under uncertainty. In this seminar, we address two challenges related to the successful application of data-driven methods: First, their performance relies heavily on the quality of the underlying dataset used. As historical data is often limited and does not contain many abnormal situations, the datasets have to be enriched through simulation. We propose a computationally efficient method to create large balanced datasets of secure and insecure operating points. Second, neural networks are treated so far as black-box tools and trained physics-agnostic, presenting a major obstacle towards their adoption in practice. For power system applications, we present a framework to provide formal guarantees of neural network behavior and to train physics-informed neural networks. We demonstrate our methods on several illustrative test cases.
Host: Deepjyoti Deka