Determining the most important variables for diffuse solar radiation prediction using adaptive neuro-fuzzy methodology; case study: City of Kerman, Iran

Identifying the most relevant variables for diffuse solar radiation prediction is of indispensable importance. In this study, the adaptive neuro-fuzzy inference system (ANFIS) is applied to select the most influential parameters for prediction of daily horizontal diffuse solar radiation (Hd). Ten important variables are nominated to analyze their effects on prediction of Hd in the city of Kerman, situated in the south central part of Iran. To achieve this, a thorough variable selection is conducted for three cases with 1, 2 and 3 inputs to introduce the best and worst inputs combinations. For the cases with 2 and 3 inputs, 45 and 120 possible combinations of inputs are considered, respectively. Providing comparisons between the most and least relevant sets of inputs reveals that appropriate selection of input parameters is an important task in prediction of Hd. For the cases with one input, it is found that sunshine duration (n) is the most influential variable. Moreover, combination of horizontal global solar radiation (H) and extraterrestrial solar radiation (Ho) as well as combination of H, Ho and n are the best sets among the cases with 2 and 3 inputs, respectively. The achieved results specify that combinations of either 2 or 3 most relevant inputs would be appropriate to provide a balance between the simplicity and high precision. Predictions using the most influential sets of 2 and 3 inputs indicate that for the ANFIS model with two inputs, the mean absolute percentage error, mean absolute bias error, root mean square error and correlation coefficient are 23.0579%, 1.0176MJ/m2, 1.3052MJ/m2 and 0.8247, respectively, and for the ANFIS model with three inputs they are 18.3143%, 0.8134MJ/m2, 1.1036MJ/m2 and 0.8783, respectively.