2023/03/03 by Ye, Feng, Brodie, Joseph, Miles, Travis +1 · 1 citation
#Applications (stat.AP) #FOS: Computer and information sciences #FOS: Electrical engineering #Systems and Control (eess.SY) #electronic engineering #information engineering
paper · doi:10.48550/arxiv.2303.02246
The reliable integration of wind energy into modern-day electricity systems heavily relies on accurate short-term wind forecasts. We propose a spatio-temporal model called AIRU-WRF (short for the AI-powered Rutgers University Weather Research & Forecasting), which combines numerical weather predictions (NWPs) with local observations in order to make wind speed forecasts that are short-term (minutes to hours ahead), and of high resolution, both spatially (site-specific) and temporally (minute-level). In contrast to purely data-driven methods, we undertake a "physics-guided" machine learning approach which captures salient physical features of the local wind field without the need to explicitly solve for those physics, including: (i) modeling wind field advection and diffusion via physically meaningful kernel functions, (ii) integrating exogenous predictors that are both meteorologically relevant and statistically significant; and (iii) linking the multi-type NWP biases to their driving mesoscale weather conditions. Tested on real-world data from the U.S. North Atlantic where several offshore wind projects are in-development, AIRU-WRF achieves notable improvements, in terms of both wind speed and power, relative to various forecasting benchmarks including physics-based, hybrid, statistical, and deep learning methods.