Comparison of BP, PSO-BP and statistical models for predicting daily global solar radiation in arid Northwest China

Abstract

Accurate prediction of global solar radiation (Rs) is important for understanding meteorological and hydrological processes, as well as the utilization of solar energy and development of clean production. In order to improve the accuracy and universality of daily Rs prediction in arid Northwest China, back-propagation neural network (BP) and BP optimized by the particle swarm optimization algorithm (PSO-BP) along with six statistical models (Ångström-Prescott, Bristow-Campbell, Swartman-Ogunlade, Sebaii, Chen and Abdalla) were adopted and compared with measured Rs data from eight representative meteorological stations across four sub-climatic zones, including the temperate continental arid zone, temperate continental high temperature-arid zone, plateau continental semi-arid zone and temperate monsoon semi-arid zone. The results showed that PSO-BP models (coefficient of determination, R2, 0.7649–0.9678) were more accurate than BP models (R2, 0.7215–0.9632) and statistical models (R2, 0.5630–0.9445) for the daily Rs prediction in the four sub-zones of arid Northwest China. The PSO-BP1 and BP1 models (with sunshine duration, maximum and minimum temperature, relative humidity and extraterrestrial radiation as inputs), PSO-BP2 and BP2 (with sunshine duration, maximum and minimum temperature and extraterrestrial radiation as inputs) performed better than the other models, with R2, mean absolute error, root mean square error, relative root mean square error and Nash-Sutcliffe coefficient ranging 0.9228–0.9678, 1.5546–1.6309 MJ·m−2·d−1, 2.0054–1.7579 MJ·m−2·d−1, 0.1517–0.1329 and 0.9017–0.9604, respectively, among which the PSO-BP1 model provided the most accurate results. Sunshine-based models (R2, 0.7533–0.9678) were generally superior to temperature-based models (R2, 0.5630–0.8492), which indicated that sunshine duration was more influential for daily Rs prediction than temperature in this area. Overall, the PSO-BP model exhibits the best generalization capability and is recommended for more accurate daily Rs prediction in arid Northwest China.