Improving generalisation capability of artificial intelligence-based solar radiation estimator models using a bio-inspired optimisation algorithm and multi-model approach

One way of reducing environmental pollution is to reduce our dependence on fossil fuels by replacing them with solar radiation (Rs), which is one of the main sources of clean and renewable energy. In this study, daily Rs values at seven meteorological stations in Iran (Ahvaz, Isfahan, Kermanshah, Mashhad, Bandar Abbas, Kerman and Tabriz) over 2010-2019 were estimated using empirical models, support vector machine (SVM), SVM coupled with cuckoo search algorithm (SVM-CSA) and multi-model approach in the form of two structures. In structure 1, data from each station were divided into training and testing sets. In structure 2, data from the former four stations were used for model training, and those from the latter three stations were used to test the models. The results showed that using meteorological parameters improved estimation accuracy compared with the use of geographical parameters for both SVM and SVM-CSA models. Coupling the CSA to SVM did improve the accuracy of radiation estimates, reducing RMSE by up to 38% (Kermanshah station) and 36% (Tabriz station) for the first structure and about 42.4% (Tabriz station) for the second. Performance analysis of the models over three intervals including, the first, middle and last third of measured radiation values at each station showed that for both structures (except at Tabriz station), the best model performance in under- and over-estimation sets of radiation values was obtained, respectively, in the first third interval (first structure, Mashhad station, RMSE = 28.39 J.cm(-2).day(-1)) and the last third interval (first structure, Bandar Abbas station, RMSE = 12.23 J.cm(-2).day(-1)). Determining the effects of climate change on Rs estimation and using remotely sensed data as inputs of the models could be considered as future works.