Seasonal Forecasting of North Atlantic Subtropical High Intensity Using a Multi-Algorithm Machine Learning Ensemble
TL;DR
It is demonstrated that a multi-algorithm machine learning ensemble can substantially improve seasonal forecasts of NASH intensity, providing a valuable reference for operational climate prediction applications.
Abstract
Accurate seasonal prediction of the North Atlantic Subtropical High (NASH) intensity is crucial for regional climate disaster preparedness, yet it remains a major challenge for current dynamical models. This study proposes a multi-algorithm ensemble forecasting system for summer NASH intensity by integrating 24 machine learning algorithms with eight selected springtime predictors. Results from the independent test period (2011–2024) show that the multi-model ensemble (MME) delivered a robust forecast, with a mean absolute error (MAE) of 0.391 hPa and a correlation coefficient (CC) of 0.898. Compared to the dynamical forecast from SEAS5 over the 1981–2024 period, the MME reduced the MAE by 40.0% and increased the CC from 0.440 to 0.816. Time Series Cross-Validation (TSCV) further confirmed the robustness of the ensemble, which outperformed most of the individual models, with an average skill improvement of 12.97%. The SHapley Additive exPlanations (SHAP) analysis applied to the Ridge model reveals that W1000ESP and VO500NEP are the two most influential predictors, and their directional contributions represent atmospheric responses to sea surface temperature anomalies, effectively capturing the cross-seasonal impacts of air–sea interactions on NASH intensity. This study demonstrates that a multi-algorithm machine learning ensemble can substantially improve seasonal forecasts of NASH intensity, providing a valuable reference for operational climate prediction applications.