Long-Term Niño3.4 index Prediction Using BiLSTM with CEEMDAN

Abstract

Traditional methods for forecasting the El Niño-Southern Oscillation (ENSO) index face significant challenges, including limited long-term forecast feasibility, substantial forecast uncertainties, and the spring predictability barrier (SPB). Recently, deep learning approaches have achieved remarkable advancements, enabling more precise and efficient ENSO forecasting results. In this study, we introduce CEEMDAN-BiLSTM, a novel hybrid modeling framework incorporating Bidirectional Long Short-Term Memory (BiLSTM) networks and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to analyze signals characterized by non-linearity and non-stationarity sequences and capture overall information of the sequence. Although most existing studies struggle to precisely predict the monthly Nino3.4 index extending up to two years horizon, our method achieves this long-term prediction. For 24-month-ahead predictions, our model has a root mean square error (RMSE) of 0.2819 and a Pearson correlation coefficient (PCC) of 0.9450, significantly outperforming alternative modeling approaches in the aspects of prediction accuracy and correlation. The method is also able to forecast the Nino3.4 index during typical strong El Niño events with two years ahead. Additionally, it achieves an average PCC of 0.8872 for 24-month-ahead spring predictions, effectively mitigating the SPB. Results also show that the CEEMDAN-BiLSTM improves both short-term (1-month-ahead PCC increased by 8.11%) and long-term (24-month-ahead RMSE decreased by 33.63%) predictions over BiLSTM. The results of the experiments indicate that CEEMDAN-BiLSTM delivers prediction results that are both more accurate and more stable and dependable in comparison to conventional neural network models.