Scope and Status of Artificial Neural Network based prediction models in Meteorology: From Backpropagation to Recurrent Networks

Artificial Neural Networks (ANNs) have, since two decades, emerged as a powerful model for forecasting. Owing to its large scale success, the backpropagation model quickly attracted the attention of meteorologists across the world for forecasting various events. ANNs are statistical models which do not take into account the dynamics of the system while making predictions. Instead, they utilize the past records for discovering the underlying statistical relation between the predictors and the predictand. Researchers have always been interested in evaluating the forecasts made by the statistical models viz-a-viz those made with the state-of-the-art Coupled Global Circulation Models (GCMs). The accuracies of both the models have been found to be comparable. Comparisons among various statistical models have also been done at length. As the backpropagation algorithm based ANN models essentially perform non-linear regression, a competition between the two approaches has been the basis of many academic studies. Of late, ANNs have been improvised to behave like a quassi-dynamical forecast model by incorporating the feedback connections, the way GCMs do. Such networks are called the Recurrent Neural Networks (RNN). The developments in such designs and their incorporation in meteorology are in a nascent stage and conclusive evidence of their accuracies over the backpropagation networks are lacking. The present article shows the role played by the ANN based computational intelligent systems in meteorology in recent years. The prospects of the dynamically modeled ANN architecture, i.e. the Recurrent Network, are discussed for application in meteorology. Whether or not such incorporation would improve the prediction accuracies is a matter of investigation. On one hand the feedback connections do provide the historical dependence but on the other hand, in meteorological dynamics, chaotic aspects may come into picture and eventually limit the predictability. The results of propsed model shows very accurate in respect of other existing models with metrics MSE, RMSE and Speed. The complementary nature of statistical and dynamical models towards each other is also outlined in this article.