(362h) Physics Informed Machine Learning for Feasibility Analysis

Using Machine Learning (ML) models to represent complex feasibility regions can allow for simpler and faster solutions to constrained optimization problems[1]. While in some applications, a feasibility region may be amorphous or dynamic (lending itself to purely data-driven approaches), many complex feasibility functions rely directly on indirectly on physical laws embedded within the constraints of the model that defines them. One such example is from the field of power-grid operations under contingency, where the N-1 secure space of an electrical grid is a complex feasible region that depends on power-flow constraints. In this work, will present a novel approach for feasibility mapping, by embedding mechanistic knowledge and structure into the latent space of Neural Network-based feasibility functions. We will show the gains obtained by employing this physics-informed ML approach, especially in sparse data regimes. Applications of this approach that will be presented include Security Constrained Optimal Power Flow [2] as well as several chemical engineering process feasibility/flexibility analysis benchmark case studies.

Acknowledment: SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525

References:

[1] Atharv Bhosekar and Marianthi Ierapetritou: Advances in surrogate based modeling, feasibility analysis and optimization: A review. Computers and Chemical Engineering, 108:250-267, 2018.

[2] Florin Capitanescu, JL Martinez Ramos, Patrick Panciatici, Daniel Kirschen, A Marano Marcolini, Ludovic Platbrood and Louis Wehenkel. State-of-the-art, challenges, and futre trends in security constrained optimal power flow. Electric Power Systems Research 81(8):1731-1741, 2011.