UQ metrics and surrogate models for plant-level wind power

DOE-funded project (through NREL), 2022–2023; no paper

Wind uncertainty
Can uncertainty metrics of a wind power model’s inputs (weather data products, forecast ensembles and turbine-level data) predict calibrated plant-level power uncertainty at wind plants that have no power measurements?

Research question

Can uncertainty metrics of a wind power model's inputs (weather data products, forecast ensembles and turbine-level data) predict calibrated plant-level power uncertainty at wind plants that have no power measurements?

My contribution

Ph.D. researcher on The University of Texas at Dallas part of the project. I developed the surrogate-modeling workflow (calibrated uncertainty targets, surrogate models and validation on held-out plants) and implemented the project's uncertainty metrics in Python, with funding from the U.S. Department of Energy through the National Renewable Energy Laboratory.

Method

Moving-window, data-source, ensemble-spread and spatial (variogram) uncertainty metrics as inputs; pinball-loss-optimal scales of a predictive distribution around the point forecast as targets at observed plants; Gaussian process regression and other regressors as surrogates; validation by holding out whole plants, scored with pinball loss, CRPS and reliability, with recalibration of the predictive distributions.

Boundary conditions

No paper. The public framework is a fresh implementation, tested on synthetic data and on a public benchmark of 264 ERCOT wind plants (ARPA-E PERFORM data, simulated plant power, 2018); project data and results are not public.