A deep learning model for two types of ENSO predictions based on GRU-TSFE and interdecadal changes in predictability

Read full text Summarise with VIRAT
Authors: Peng Dong, Chundi Hu, Guanchao Tong, Zeming Wu, Chengjie Lei
Cited by: 0
References: 60

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

ENSO drives global climate variability and extremes, motivating skillful mid-range forecasts of ENSO indices. Here, we propose a three-branch deep-learning framework, the Seasonal Gated Recurrent Unit with Time-Series Feature Extraction (GRU-TSFE), which integrates convolutional feature extraction with recurrent temporal modeling for continuous monthly-lead prediction. Twelve-month hindcasts of Niño3, Niño4, NCT, and NWP indices (March 1985–February 2025) show comparable skill at 1–4-month leads, but substantially higher accuracy for Niño4 and NWP at 5–12-month leads. The GRU-TSFE model performs best for Niño4, maintaining correlation coefficients near 0.80 up to 10-month lead times. Comparison of prediction results from two consecutive 20-year periods reveals an interdecadal migration of the spring predictability barrier, with a delayed onset for Niño3 and an earlier onset for NCT. These shifts may be associated with PDO-related background variability, tropical Pacific mean-state changes, and evolving ENSO modal structure; these mechanisms are treated here as plausible contributors and testable hypotheses rather than established causal drivers.

Fields of Study & Subjects

Climate variability and models Meteorological Phenomena and Simulations Hydrological Forecasting Using AI

Publication details

Journal / source
Artificial Intelligence in Geosciences
Volume
7
Issue
4
Pages
100279–100279
Publication date
29 September 2026
Language
EN