Learning Frequency Nadir From Multi-Fidelity Data For Dynamic Secure Microgrid Islanding

A frequency-constrained microgrid scheduling model has been used to obtain dispatch commands that ensure dynamic frequency security under event-triggered islanding. Among all frequency constraint encoding approaches, data-driven methods show superiority in terms of capturing more generic frequency behavior but are limited by the quality of data. In this paper, we introduce a deep learning method for frequency nadir prediction that can be trained using multi-fidelity data. Without loss of generality, we consider the training dataset consists of large-size low-fidelity data and small-size high-fidelity data. Instead of directly performing the training over either dataset, we first learn the correlation between the low- and high-fidelity dataset. Then, this correlation model can be used to generate a large size of synthetic high-fidelity data with negligible computation effort. The multi-fidelity training admits a 95% error reduction in out-of-sample testing. Once being encoded into the microgrid scheduling model, less conservative dispatch commands can be obtained.