Prediction of Global Solar Radiation Using Artificial Neural Network Model for Coastal District of Karnataka

Solar energy has enormous direct and indirect applications. The direct ones are solar water heating, crop drying and the indirect ones are probably electricity generation using photovoltaic technology, heliostatic powerplants, hydrogen generation using solar power electrolyser are the few among many other applications. The prediction of daily global solar radiation (GSR) data is important for many of these solar applications, and other applications on renewable energy that can be found in meteorological studies, for small to long range of data accumulation. This work aims at Prediction of global solar radiation for a particular place in India which bases itself on several input parameters using Artificial Neural Networks (ANN). The study is carried out in Nitte, Udupi which is a village in Karkala taluk of Udupi district, in the coastal district state of Karnataka having latitude 13.100N and longitude 74.930E and is at an altitude of about 265 feet above sea level.  The empirical models available to estimate global solar radiation for dedicated places the accuracies have been found to be low which in fact is the requirement for a dedicated model to be provable in the later stages when there is variation within the parameters. Artificial neural networks it has been found in many cases to give better prediction accuracies. The neural network model made in this work has been built using the neural network toolbox in Mat lab version 7.13. Artificial Neural Network modelling will be done using Multilayer Perceptron Neural Network model.

MLP model has been considered for the creating the ANN model which is done using the measured data, where the input parameters are decided to be air temperature, relative humidity, time of the day, wind velocity, wet bulb temperature, atmospheric pressure, sunshine hours, solar angle, clearance index, declination angle and the global solar radiation has been the output parameter. The collected data from measurements has been divided into two parts, the first part will used for training and the latter part will be used for testing the created neural network model. Training data set, which contains 85% of the data and test data set, which contains 15 % of the data, selected randomly.  The suitable number of hidden layer has been selected, such that the overall accuracy is maximized. In order to select the best training algorithm, the MLP model has been trained using different training algorithms like trainscg, trainrp, trainlm, trainbfg and trainoss. The evolved MLP model from the entire training and testing has been able to predict the estimated and experimental values of global solar irradiation with an accuracy of 86.69% for the test data. From the overall results it has been found there is good agreement between the predicated using ANN and experimental values of global solar irradiation.