Spatio-temporal modelling and uncertainty estimation of hourly global solar irradiance using Extreme Learning Machines

Abstract

Abstract Solar photovoltaic (PV) is one of the most promising technologies for the transition from fossil fuels to renewable energy production. Accurate spatial and temporal modelling of solar irradiance is a key factor in the evaluation of PV technology potential for harvesting solar energy. We present here a data-driven approach based on an ensemble of Extreme Learning Machines using geographic and topographic features in input to predict the global horizontal irradiance in Switzerland from coarse-resolution satellite measurements. This provides a precise mapping of hourly global solar irradiance for each (250 × 250) m2 pixel of a grid covering the entire country. The uncertainty on predicted values is quantified through a variance-based analysis, able to distinguish between model and data uncertainty. The former amounts to 1%, whereas the latter is close to 15% of the predicted values. The presented methodology is scalable and applicable to any large environmental dataset. Our modelling of solar irradiance at hourly temporal resolution and of its uncertainty will allow for an estimate of hourly PV potential in Switzerland to facilitate a more efficient integration of solar photovoltaics into the built environment.