Grouped Gaussian Processes for Solar Forecasting

Astrid Dahl, Edwin V Bonilla

Abstract: We propose a Gaussian process regression network model where latent functions are arbitrarily coupled a priori. Driven by the problem of developing forecast methods for distributed solar and other renewable power generation, we propose coupled priors that exploit spatial dependencies in a scalable structure. We estimate short term forecast models for solar power at multiple distributed sites and ground wind speed at proximate weather stations. Our approach maintains or improves point-forecast accuracy relative to competing benchmarks. At the same time our approach significantly reduces forecast variance.

TL;DR: We develop scalable multi-task Gaussian Process models with coupled priors that exploit spatial dependence for distributed solar forecasting.

Keywords: Gaussian processes, multi-task learning, nonparametric Bayesian methods, scalable inference