Data fusion of complementary data sources using Machine Learning enables higher accuracy Solar Resource Maps

TL;DR

This paper demonstrates that machine learning-based data fusion of multiple solar and atmospheric data sources significantly improves the accuracy of solar resource maps, exemplified by a 30-year GHI map over Norway.

Contribution

The study introduces a machine learning data fusion approach that combines diverse data sources to enhance solar resource mapping accuracy, applicable globally.

Findings

ML fusion improves GHI estimate accuracy over individual data sources

Generated a 30-year high-resolution GHI map for Norway

Open data release of the ML-corrected solar resource map

Abstract

In the present work, we collect solar irradiance and atmospheric condition data from several products, obtained from both numerical models (ERA5 and NORA3) and satellite observations (CMSAF-SARAH3). We then train simple supervised Machine Learning (ML) data fusion models, using these products as predictors and direct in-situ Global Horizontal Irradiance (GHI) measurements over Norway as ground-truth. We show that combining these products by applying our trained ML models provides a GHI estimate that is significantly more accurate than that obtained from any product taken individually. Using the trained models, we generate a 30-year ML-corrected map of GHI over Norway, which we release as a new open data product. Our ML-based data fusion methodology could be applied, after suitable training and input data selection, to any geographic area on Earth.