Comparison of Neural Network Training Functions for Prediction of Outgoing Longwave Radiation over the Bay of Bengal

Abstract

Artificial Neural Network (ANN) is contributed in many fields for forecasting or prediction like medical field, Sensex, biometric, remote sensing, pattern recognition, etc. The beauty of ANN is that it is having a capability to dealing with non-linear data. In the present situation, everyone is facing atmospheric and climate change problems. Study on an atmospheric and meteorological data is a challenging task because it gives a complex and non-linear type of data. In such a case, ANN is a better technique for dealing with non-linear atmospheric and meteorological data. In the present analysis, use an atmospheric data such as wind speed, latent heat flux, net surface heat flux and Sea Surface Temperature for the prediction of outgoing longwave radiation over the latitude 80°E:100°E and longitude 0–25°N for the Bay of Bengal. To achieve the objective, use a Feedforward Neural Network with five different training functions like Levenberg–Marquardt, conjugate gradient with Beale–Powell restarts, one step secant, Broyden–Fletcher–Goldfarb–Shanno (BFGS) quasi-Newton and gradient descent. Use various performance analysis parameters such as correlation coefficient, root mean square error and prediction accuracy to find out the performance of training function. From the analysis, it is observed that from the above five training functions, Levenberg–Marquardt gives a 95.81% accuracy which is higher than the others.