Machine learning models to quantify and map daily global solar radiation and photovoltaic power

Global solar radiation (R-s) reaching Earth's surface is the primary information for the design and application of solar energy-related systems. High-resolution R-s measurements are limited owing to the high costs of measuring devices, and their stringent operational maintenance procedures. This study evaluated a newly developed machine learning model, namely the hybrid particle swarm optimization and extreme learning machine (PSO-ELM), to accurately predict daily R-s. The newly proposed model was compared with five other machine learning models, namely the original ELM, support vector machine, generalized regression neural networks, M5 model tree, and autoencoder, under two training scenarios using long-term R-s and other climatic data taken during 1961-2016 from seven stations located on the Loess Plateau of China. Overall, the PSO-ELM with full climatic data as inputs provided more accurate R-s estimations. We also calculated the daily R-s at fifty other stations without R-s measurements on the Loess Plateau using the PSO-ELM model, as well as the potential photovoltaic (PV) power using an empirical PV power model, and then generated high-resolution (0.25 degrees) R-s and PV power data to investigate the patterns of R-s and PV power. Significant reductions in R-s (- 6.49 MJ m(-2) per year, p < 0.05) and PV power (- 0.46 kWh m(-2) per year, p < 0.05) were observed. The northwestern parts of the study area exhibited more R-s and PV power and are therefore considered more favorable for solar energy-related applications. Our study confirms the effectiveness of the PSO-ELM for solar energy modeling, particularly in areas where in-situ measurements are unavailable.