Machine learning prediction of the Madden–Julian oscillation using reservoir computing

Abstract

The prediction of the Madden–Julian Oscillation (MJO), a massive tropical weather event with global socio-economic impacts, has been infamously difficult with physics-based weather prediction models. We employ the reservoir computing, a brain-inspired machine-learning technique, to construct a machine learning model that forecasts the real-time multivariate MJO index (RMM), a macroscopic variable that represents the state of the MJO. The training data was refined by development of a novel real-time band-pass filter that extracts the recurrency of MJO signals only from the past raw atmospheric data, and by selection of a suitable time-delay coordinate of the RMM that enhances the recurrency of the input data. The constructed model demonstrated the skill to forecast the time sequence of the RMM for a month from pre-developmental stages of the MJO. Examination of best-performing cases suggested that RMM sequences may be predicted for over two months in some cases. These results imply that inherent predictability limit of the MJO is longer than that has been estimated from physics-based weather prediction models.