Development of Deep Convolutional Neural Network Ensemble Models for 36-Month ENSO Forecasts

Abstract

The state of the El Niño-Southern Oscillation (ENSO) has chaotic yet deterministic seasonal patterns and inter-annual fluctuations over the equatorial Pacific Ocean. ENSO has impacts and global teleconnections on regional temperature, precipitation, and mid-tropospheric atmospheric circulation and has been used as a predictor of regional weather. Despite being developed over several decades, dynamical and statistical models are still unable to reliably predict seasonal ENSO. This paper presents the unique utilization of several deep convolutional neural networks, identified preferable model parameters, and an optimized ensemble output to extend the ENSO forecast by up to 36 months in advance. While individual models performed differently depending on the forecasting lead month, the ensemble output is the only model that produces a correlation of 0.52 with an index of agreement of 0.60 for the 36th month forecast, a 4% and 7% improvement in the cumulative index of agreement and r score, respectively, over the best single model. The results demonstrate the moderate ENSO forecasting capability of the system and the next step in extending the prediction lead time to previous generations of ENSO forecasting models.