Entropy criterion for surrogate timeseries data generation via non-parametric dimensionality reduction

Surrogate data are necessary for increasing data availability for data-intensive engineering analyses, e.g., optimization, by generating artificial data instances that preserve both the trends as well as the randomness inherent in the raw data, allowing the analyst to expand the usable data for downstream analyses. This manuscript focuses on the generation of surrogate timeseries data for applications within the HERON framework developed by Idaho National Laboratory to optimize resource allocation of a nuclear reactor with cogeneration capabilities, i.e., steam or electricity production. The HERON surrogate data generation relies on a user-defined Fourier-based algorithm for detrending seasonal behavior and an autoregressive moving-average (ARMA) algorithm for preserving the statistical nature of the detrended data. A key limitation of this algorithm is that the resultant errors may not be normally distributed, thus reducing confidence in the statistical consistency between the raw and surrogate data. To overcome this limitation, this manuscript proposes an alternative data-driven non-parametric algorithm whose dimensionality reduction is determined by an entropy-based cutoff criterion to hedge against overfitting and ensure statistical consistency. This manuscript develops the proposed algorithm, called NEST, and compares it to HERON surrogate data using several quantitative tests.