Long-term ENSO prediction with echo-state networks

Abstract The El Niño-Southern Oscillation (ENSO) is a climate phenomenon that profoundly impacts weather patterns and extreme events worldwide. Here we develop method based on recurrent neural network, called echo state network (ESN), which can be trained efficiently to predict different ENSO indices despite their relatively high noise levels. To achieve this, train the ESN model low-frequency variability of estimate potential future high-frequency from specific samples its past history. Our reveals importance cross-scale interactions in mechanisms underlying skilfully predicts especially Niño at lead times up 21 months. This study considers forecasts skillful if correlation coefficients are above 0.5. results show component carries substantial predictive power, exploited by training our single scalar time series. proposed machine learning for data-driven modeling readily applied other series, e.g. finance physiology. However, it should noted approach cannot straightforwardly turned into real-time operational forecast because decomposition original series slow fast components using low-pass filter techniques.