Reconstruction of a dynamical-statistical forecasting model of the ENSO index based on the improved self-memorization principle

To address the inaccuracy of long-term El Niño-Southern Oscillation (ENSO) forecasts, a new dynamical-statistical forecasting model of the ENSO index was developed based on dynamical model reconstruction and improved self-memorization. To overcome the problem of single initial prediction values, the largest Lyapunov exponent was introduced to improve the traditional self-memorization function, thereby making it more effective for describing chaotic systems, such as ENSO. Equation reconstruction, based on actual data, was used as a dynamical core to overcome the problem of using a simple core. The developed dynamical-statistical forecasting model of the ENSO index is used to predict the sea surface temperature anomaly in the equatorial eastern Pacific and El Niño/La Niña events. The real-time predictive skills of the improved model were tested. The results show that our model predicted well within lead times of 12 months. Compared with six mature models, both temporal correlation and root mean square error of the improved model are slightly worse than those of the European Centre for Medium-Range Weather Forecasts model, but better than those of the other five models. Additionally, the margin between the forecast results in summer and those in winter is not great, which means that the improved model can overcome the “spring predictability barrier”, to some extent. Finally, a real-time prediction experiment is carried out beginning in September 2014. Our model is a new exploration of the ENSO forecasting method.