Real-time localised forecasting of the Madden-Julian Oscillation using neural network models

Existing statistical forecast models of the Madden-Julian Oscillation (MJO) are generally of very low order and predict the evolution of a small number (typically two) of principal components (PCs). While such models are skilful up to 25 days lead time, by design they only predict the very largest-scale features of the MJO. Here we present a higher-order MJO statistical forecast model that is able to predict MJO variability on smaller, more localised scales, that will be of more direct benefit to national weather agencies and regional government planning. The model is based on daily outgoing long-wave radiation (OLR) data that are intraseasonally filtered using a recently developed technique of empirical mode decomposition that can be used in real time. A standard truncated PC analysis is then used to isolate the maximum amount of variance in a finite number of modes. The evolution of these modes is then forecast using a neural network model, which does not suffer from the parametrisation problems of other statistical forecast techniques when applied to a higher number of modes. Compared to a standard 2-PC model, the higher-order PC model showed improved skill over the whole MJO domain, with substantial improvements over the western Pacific, Arabian Sea, Bay of Bengal, South China Sea and Phillipine Sea.