Spatially and temporally explicit model to predict solar radiation using machine learning

Abstract

Accurate prediction of solar radiation can provide numerous benefits to society and individuals. Historically, these predictions have been provided by Numerical Weather Prediction (NWP) models, satellite extrapolation, and other methods, with machine learning (ML) techniques being a recent addition. The study that follows examines three ML techniques performed using data across the state of Delaware. Using local, high-resolution mesonet data, coupled with satellite observations from the operational geostationary satellite (GOES-R), for the period 2018 – 2021, a model for each of the three techniques is developed. Each of these techniques will be employed to predict future solar radiation measurements and gap-fill current observations based on prevailing meteorological conditions. ☐ ML model performance is investigated using common statistical criteria (Mean Absolute Error, Root Mean Squared Error, and R-squared) to identify an optimal ML model technique for the study area. Techniques to reduce required model inputs are utilized to simplify generated models while maintaining predictive accuracy. Based on these simplification methods, air temperature, relative humidity, and the geostationary satellite data are identified as primary meteorological drivers of model accuracy. The results of the study show ML methods can be used to make accurate predictions of solar radiation over the study area, both temporally and spatially. XGboost performs best of the ML methods employed, followed by random forest, and linear regression.