Forecast El Niño – southern oscillation (ENSO), using artificial neural networks

Abstract

This study presents an approach based on artificial neural networks for forecasting the El Niño–Southern Oscillation (ENSO) phenomenon, using the Oceanic Niño Index (ONI) as the target variable. Four neural network architectures were evaluated — Multilayer Perceptron (MLP), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and CNN-LSTM — across forecasting horizons defined by temporal lags ranging from 1 to 6 months. The main objective was to analyze the impact of lead time on prediction accuracy and to determine the practical time limit of these estimates. The predictor variables include well-established oceanographic and atmospheric indicators: Sea Surface Temperature (SST) anomalies in the Niño 3.4 region, the Southern Oscillation Index (SOI), the Multivariate El Niño Index version 2 (MEI.v2), and the Pacific North American Pattern (PNA). To prevent data leakage, a temporal split was applied between the training and testing sets, without shuffling. The results indicate that, for lag 1, the models achieved coefficients of determination (R²) above 0.94, demonstrating high short-term forecasting capability. However, as the lag increased, a progressive decline in performance was observed, reaching R² values below 0.20 for lag 6. A comparative analysis using a unified score based on normalized R², MAE, and RMSE revealed the robustness of MLP across different temporal windows, although LSTM outperformed in longer-term horizons. These results highlight the potential of neural networks in capturing short-term climate patterns, while also emphasizing their significant limitations for extended forecasting timeframes.