Deep Neural Networks for ENSO Prediction in the Niño 3.4 and Niño 1+2 Regions

Forecasting the El Niño-Southern Oscillation (ENSO) is one of the most challenging tasks because ENSO events show large differences in their amplitude, temporal evolution, and spatial pattern. This research discusses a deep neural network model that integrates convolutional short-term and long-term memory (ConvLSTM) to forecast the patterns of sea surface temperature that represent the evolution of El Niño - Southern Oscillation (ENSO) over different time horizons with simultaneous space-time sequences in the Niño 3.4 and Niño 1+2 regions. Monthly gridded Sea Surface Temperature data and Anomaly Index data were used in El Niño regions for the period 1854-2021. In general, the methodology includes the Sea Surface Temperature (SST) normalization of the tropical Pacific, the creation and selection of the architecture of the deep neural network model, the training and validation. Furthermore, the forecast and evaluation of the SST and anomalies in space-time with the El Niño of 1982/1983 and 1997/98 and with different time horizons for the test stage were developed. The results show a good model performance with ENSO data for El Niño 1982/83 (training) and El Niño 1997/98 (validation) in space and time, and with climatic indices for regions 3.4 and 1+2. The model performance was evaluated using statistical metrics of observed and forecasted data over six months in the equatorial Pacific Ocean (latitude 10°S–10°N and longitude 140°E–80°W). Likewise, a better forecast is obtained in the Niño 3.4 region than in the Niño 1+2 region; with very good results in the first three months, reducing its prediction efficiency as the forecast month’s increase.