Deep learning with autoencoders and LSTM for ENSO forecasting

Abstract

El Niño Southern Oscillation (ENSO) is the prominent recurrent climatic pattern in the tropical Pacific Ocean with global impacts on regional climates. This study utilizes deep learning to predict the Niño 3.4 index by encoding non-linear sea surface temperature patterns in the tropical Pacific using an autoencoder neural network. The resulting encoded patterns identify crucial centers of action in the Pacific that serve as predictors of the ENSO mode. These patterns are utilized as predictors for forecasting the Niño 3.4 index with a lead time of at least 6 months using the Long Short-Term Memory (LSTM) deep learning model. The analysis uncovers multiple non-linear dipole patterns in the tropical Pacific, with anomalies that are both regionalized and latitudinally oriented that should support a single inter-tropical convergence zone for modeling efforts. Leveraging these encoded patterns as predictors, the LSTM - trained on monthly data from 1950 to 2007 and tested from 2008 to 2022 - shows fidelity in predicting the Niño 3.4 index. The encoded patterns captured the annual cycle of ENSO with a 0.94 correlation between the actual and predicted Niño 3.4 index for lag 12 and 0.91 for lags 6 and 18. Additionally, the 6-month lag predictions excel in detecting extreme ENSO events, achieving an 85% hit rate, outperforming the 70% hit rate at lag 12 and 55% hit rate at lag 18. The prediction accuracy peaks from November to March, with correlations ranging from 0.94 to 0.96. The average correlations in the boreal spring were as large as 0.84, indicating the method has the capability to decrease the spring predictability barrier.

Comparison of machine learning models in forecasting different ENSO types

“…Unlike models that may struggle with the temporal patterns in a time series, FFNN, and LSTM can capture and leverage the underlying dependencies present in the past six months' data to make accurate predictions for the upcoming six months. Recently, consistent with our results, several other studies have successfully applied neural networks to forecast ENSO with promising accuracy at several lead times, and at times outperforming dynamical models [11,17,[45][46][47][48][49][50][51][52]. For example, the study by Ham et al [47] applied deep learning with a convolutional neural network (CNN) to forecast the Niño 3.4 index in up to 17 months lead time, achieving a correlation of at least 0.5.…”

“…Hu et al [45] applied deep residual convolutional neural network and heterogeneous transfer learning to achieve 83.3% accuracy for forecasting EI Niño type in 12-month lead time. Similar impressive results were also achieved by other studies applying deep learning for ENSO forecasting [11,50].…”

The influence of global warming on ENSO dynamics

“…In prior work, we have demonstrated that the Nonhomogeneous Hidden Markov Model (NHMM) applied to gridded, monthly SST data can objectively identify 5 ENSO flavors (whose spatial patterns and time of occurrence are similar to those of the 5 types of events typically identified), their temporal dynamics 24 , and their implications for global precipitation 23 . Further its 1-18 month ahead prediction skill for the benchmark Niño 3.4 index is comparable to that of other popular machine learning models 19,[25][26][27] . Given its interpretability and relative accuracy, we use it to explore how the temporal dynamics of ENSO may have changed relative to changing global temperature.…”

“…It is well recognized that using a single ENSO index does not adequately represent the spatiotemporal dynamics or allow for an effective long-term ENSO prediction model 14,1516,17 . In addition to deterministic physics-based models, a variety of machine learning methods [18][19][20][21][22][23] have been advanced to simulate the dynamics of ENSO diversity. In prior work, we have demonstrated that the Nonhomogeneous Hidden Markov Model (NHMM) applied to gridded, monthly SST data can objectively identify 5 ENSO flavors (whose spatial patterns and time of occurrence are similar to those of the 5 types of events typically identified), their temporal dynamics 24 , and their implications for global precipitation 23 .…”