موسسه ژئوفیزیک دانشگاه تهرانفیزیک زمین و فضا2538-371X49420240220Regression Models and Hybrid Intelligent Systems for Estimating Clear-Sky Downward Longwave Radiation in Equatorial AfricaRegression Models and Hybrid Intelligent Systems for Estimating Clear-Sky Downward Longwave Radiation in Equatorial Africa1431599207110.22059/jesphys.2023.352877.1007484FANsikan Ime ObotDepartment of Physics, Faculty of Science, University of Lagos, Akoka, Lagos, Nigeria.0000-0003-2289-3711Journal Article20221225Modelling downward longwave radiation (DLR) in Equatorial Africa is challenging due to dense cloud cover and data scarcity. In this twofold study, daily cloudless DLR in Ilorin (8° 32′ N, 4° 34′ E), Nigeria, was modelled using two atmospheric factors, namely water vapour pressure and air temperature. Firstly, four cloudless DLR models were reformed and tested with others. Secondly, both particle swarm optimization (PSO) and genetic algorithm (GA) were deployed to optimize the adaptive neuro-fuzzy inference system (ANFIS) and artificial neural networks (ANN). The statistical measures used to evaluate the performance of the models were the coefficient of determination (<em>r<sup>2</sup></em>), the mean bias error, and the mean square error (<em>MSE</em>). While restructuring clear skies DLR models typically reduces the estimation errors, it may not necessarily impact <em>r<sup>2</sup></em> positively. The regression models have <em>r<sup>2</sup></em> values ranging from approximately 0.82 to 0.87, while <em>MSE</em> lies between 56.6 W/m<sup>2</sup> and 767.5 W/m<sup>2</sup>. There are instances where <em>MSE</em> drastically reduces from 692.6 to 72.3 (W/m<sup>2</sup>) and from 767.5 to 66.2 (W/m<sup>2</sup>) after restructuring two different models. A recently developed expression for the region remains the best, possibly because of its format. During the training phase of the computationally intelligent systems, <em>r<sup>2</sup></em> approximately ranges between 88% and 92% but lies between 55% and 76% during testing. Although reproducibility inclusion in the code can meaningfully improve ANN systems at training, GA optimizes better than PSO. Furthermore, hybrid intelligent systems had higher <em>r<sup>2</sup></em> values than standalone computationally intelligent modes at the testing phase. Due to the efficient generalization based on <em>r<sup>2</sup></em> during the testing phase, ANN-GA is viable for modelling cloudless DLR at this site, though ANFIS has the lowest <em>MSE</em> at this same stage.Modelling downward longwave radiation (DLR) in Equatorial Africa is challenging due to dense cloud cover and data scarcity. In this twofold study, daily cloudless DLR in Ilorin (8° 32′ N, 4° 34′ E), Nigeria, was modelled using two atmospheric factors, namely water vapour pressure and air temperature. Firstly, four cloudless DLR models were reformed and tested with others. Secondly, both particle swarm optimization (PSO) and genetic algorithm (GA) were deployed to optimize the adaptive neuro-fuzzy inference system (ANFIS) and artificial neural networks (ANN). The statistical measures used to evaluate the performance of the models were the coefficient of determination (<em>r<sup>2</sup></em>), the mean bias error, and the mean square error (<em>MSE</em>). While restructuring clear skies DLR models typically reduces the estimation errors, it may not necessarily impact <em>r<sup>2</sup></em> positively. The regression models have <em>r<sup>2</sup></em> values ranging from approximately 0.82 to 0.87, while <em>MSE</em> lies between 56.6 W/m<sup>2</sup> and 767.5 W/m<sup>2</sup>. There are instances where <em>MSE</em> drastically reduces from 692.6 to 72.3 (W/m<sup>2</sup>) and from 767.5 to 66.2 (W/m<sup>2</sup>) after restructuring two different models. A recently developed expression for the region remains the best, possibly because of its format. During the training phase of the computationally intelligent systems, <em>r<sup>2</sup></em> approximately ranges between 88% and 92% but lies between 55% and 76% during testing. Although reproducibility inclusion in the code can meaningfully improve ANN systems at training, GA optimizes better than PSO. Furthermore, hybrid intelligent systems had higher <em>r<sup>2</sup></em> values than standalone computationally intelligent modes at the testing phase. Due to the efficient generalization based on <em>r<sup>2</sup></em> during the testing phase, ANN-GA is viable for modelling cloudless DLR at this site, though ANFIS has the lowest <em>MSE</em> at this same stage.https://jesphys.ut.ac.ir/article_92071_6e5f3c4c76ca514f7d14b4102001fa44.pdf