Enabling predictive reduced order modeling of high-fidelity wind plant simulations with in-situ modal decomposition and basis interpolation

As wind plant simulation capabilities approach exascale, new challenges emerge regarding reduced order models (ROM) for realtime controls, uncertainty quantification, and data compression. Exascale paradigms favor ROM techniques that minimize storage and communication in a distributed environment. Many reduced order modeling and machine learning techniques require a decomposition of a snapshot matrix that is prohibitively expensive to store and access in exascale simulations. To overcome this barrier, we demonstrate a single-pass randomized SVD of high fidelity wind plant simulations that minimizes storage and communication requirements. These matrix factorizations are used to develop a linear parameter-varying dynamic mode decomposition model that can smoothly interpolate the reduced order model to unseen inflows. A key challenge for these systems is determining linearizations at new operating points. We compare two possible approaches for obtaining new linearizations: local basis interpolation on Stieffel manifolds, and using a lower fidelity analytical model. These developments enable truly predictive ROM’s for exascale simulation capabilities.