Neural networks for short-term climate prediction

There are 2 main paradigms for short-term climate prediction: (1) General Circulation Models (GCMs) and (2) linear statistical methods (CCA, SVD). The advent of the neural network (NN) model, a nonlinear empirical technique developed originally in the field of artificial intelligence (Al), offers the hope that NN may improve the forecast skills attained by linear statistical models. However, the application of NN to climate prediction has not been straightward. The main reasons why NN has difficulty being adapted for short-term climate prediction ate: (a) the NN problem is ill-conditioned with a relatively short climate data record, and (b) in climate studies, large spatial data fields have to be dealt with. For (a), ensemble averaging is found to be effective, and for (b), the empirical orthogonal function (EOF) method is useful. Better ways of visualizing the neural networks results are also devised. The method is applied to El Ni (n) over tilde o forecasting.