Machine Learning Methods for ENSO Analysis and Prediction

The El Niño-Southern Oscillation (ENSO) plays a vital role in the interannual variability of the global climate. In order to reduce its adverse impacts on society, many statistical and dynamical models have been used to predict its future states. However, most of these models present a limited forecast skill for lead times beyond 6 months. In this paper, we present and discuss results from previous work and describe the University of Brasilia/Columbia Water Center (UNB/CWC) ENSO forecast model, which has been recently developed and incorporated into the ENSO Prediction Plume provided by the International Research Institute for Climate and Society. The model is based on a nonlinear method of dimensionality reduction and on a regularized least squares regression. This model is shown to have a skill similar to or better than other ENSO forecast models, particularly for longer lead times. Many dynamical and statistical models predicted a strong El Niño event in 2014. The UNB/CWC model did not, consistent with the subsequent observations. The model’s ENSO predictions for 2014 are presented and discussed.