Solar irradiance is an ultimate source of solar energy and it depends on various parameters including temperature, humidity, wind speed etc.Therefore, prediction of solar irradiance will help in planning and designing efficient solar energy system. Many research groups have attempted to predict the solar irradiance using other geographical parameter with the help of machine learning algorithms. Out of them artificial neural networks comes out to be the most common. In the present research work a variety of neural network architectures are implemented to find out the best and optimum method to predict the solar irradiance.In addition, all the other parameters such as training function (trainlm, trainbr, and trainscg).Present research work shows an implementation of artificial neural network, recurrent neural network, general regression neural network and radial basis function network fully crossed with three training functions (trainbr, trainscg, and trainlm). The output is judged on three other metrics apart from the correlation between original and predicted solar irradiance. Thesemetrics consists of mean bias error(MBE),mean absolute error (MAE), root mean square error (RMSE). All the methods are implemented on dataset obtained from national solar radiation database (NSRDB). The dataset contain eight other attributes apart from solar irradiance for eight years (2000-2007). The prediction of global solar irradiance is done by taking all other eight attributes as input. Proposedmethods are also implemented on different ratios of training and testing data ranging from 50% to 90% in training and 50% to 10 % in testing. Further, implementation of principal componentanalysis is also done to reduce the redundancy between attributes.Overall, in all the methods general regression neuralnetworkcomes out to be the most efficient one by producing99.44% correlation coefficient between predicted and original global solar irradiance. This showsbetter performanceof general regression neural network overcommon artificial neural network. The error terms are also minimumin the case of general regression neural network.