Analysis of different combinations of meteorological parameters in predicting the horizontal global solar radiation with ANN approach: A case study

Solar energy is a clean renewable energy source and availability of solar resources at a particular location depends on the local meteorological parameters. In the present study, prediction models using artificial neural networks (ANN) are developed by varying the meteorological parameters from one to six. A two year database of daily global solar radiation (GSR), daily minimum temperature (T-min), daily maximum temperature (T-max), difference of daily maximum and minimum temperature (DT), sunshine hours (S), theoretical sunshine hours (S-o) and extraterrestrial radiation (H-o) have been used to train the ANN. Six ANN models are developed (ANN-1 to ANN-6) with 32 possible combinations of inputs and are used to train the network to identify the best combination of inputs to estimate the monthly mean daily GSR accurately. All the models are validated and the performance of the models are analyzed by using the statistical tools. Out of the six ANN models with all the possible combination of input variables, ANN-2 and ANN-3 have given best prediction with the combinations of [DT, H-o] and [DT, H-o, S-o] respectively. The statistical tool Relative Root Mean Square Error (RRMSE) showed the least value of 3.96% with [DT, H-o] inputs. The ANN-1 trained with calculated approximate sunshine hours (S-a) has also shown high prediction accuracy. Sunshine based model and temperature based model are validated with ANN-1, ANN-2 and ANN-3 architectures. Results showed that the developed ANN models outperform the considered empirical models. The combination of [S-o, H-o] has produced excellent estimation, which are theoretical parameters and does not require any measured meteorological parameters. Superior performance is observed with less number of inputs which are readily available for any location.