Theory of spatiotemporal deep analogs and their application to solar forecasting

Weiming Hu, Guido Cervone, George Young

Research output: Chapter in Book/Report/Conference proceedingChapter

Abstract

Analog forecasting is a method widely used for generating future forecasts by matching current forecast with historical weather patterns. Under optimal conditions, it was shown to generate accurate and well-calibrated forecasts. However, under suboptimal conditions, which include extreme or rare weather patterns, a small search repository, or a changing relationship between forecasts and observed outcomes, analogs can fail and do so without a known level of uncertainty. The Analog Ensemble technique was developed to address these shortcomings. Recently, it was shown that when the search for similar weather analogs is extended to the spatial domain, better analogs can be found because they can cope with the spatial and temporal autocorrelation of weather variables. However, including spatial information in the selection of analogs is not a trivial task. This chapter introduces an analog ensemble method that uses convolutional neural networks to identify optimal analogs. Convolutional neural networks are effective at extracting high-level spatial features. The proposed network is tested on solar irradiance forecasting, an important variable for renewable energy generation with a close tie to atmospheric conditions. It measures the amount of energy received from the Sun in the form of electromagnetic radiation. Results show that the introduction of spatial information into the similarity metric leads to about a 10% decrease in prediction error. The similarity metric also demonstrates spatial robustness when nearby forecasts are searched. The efficacy of the trained model has been studied and verified with various interpretation methods and visualization.

Original languageEnglish (US)
Title of host publicationArtificial Intelligence in Earth Science
Subtitle of host publicationBest Practices and Fundamental Challenges
PublisherElsevier
Pages205-246
Number of pages42
ISBN (Electronic)9780323917377
ISBN (Print)9780323972161
DOIs
StatePublished - Jan 1 2023

All Science Journal Classification (ASJC) codes

  • Earth and Planetary Sciences(all)

Cite this