Deep Learning Meets Teleconnections: Improving S2S Predictions for European Winter Weather

TL;DR

This study addresses subseasonal-to-seasonal forecasting of European winter weather by embedding physical teleconnections into deep learning models. It compares a plain LSTM, an Index-LSTM that augments inputs with SPV and MJO indices, and a ViT-LSTM that encodes spatiotemporal fields (u10 and OLR) via Vision Transformers with MAE pretraining. Results show teleconnection-aware models improve long-range skill, with ViT-LSTM outperforming ECMWF hindcasts beyond $4$ weeks for several NAE regimes and offering insight into precursor patterns and teleconnection pathways. The work demonstrates that physics-guided DL can both enhance predictive skill and provide new mechanistic understanding of atmospheric dynamics, supporting more reliable and interpretable S2S forecasts.

Abstract

Predictions on subseasonal-to-seasonal (S2S) timescales--ranging from two weeks to two month--are crucial for early warning systems but remain challenging owing to chaos in the climate system. Teleconnections, such as the stratospheric polar vortex (SPV) and Madden-Julian Oscillation (MJO), offer windows of enhanced predictability, however, their complex interactions remain underutilized in operational forecasting. Here, we developed and evaluated deep learning architectures to predict North Atlantic-European (NAE) weather regimes, systematically assessing the role of remote drivers in improving S2S forecast skill of deep learning models. We implemented (1) a Long Short-term Memory (LSTM) network predicting the NAE regimes of the next six weeks based on previous regimes, (2) an Index-LSTM incorporating SPV and MJO indices, and (3) a ViT-LSTM using a Vision Transformer to directly encode stratospheric wind and tropical outgoing longwave radiation fields. These models are compared with operational hindcasts as well as other AI models. Our results show that leveraging teleconnection information enhances skill at longer lead times. Notably, the ViT-LSTM outperforms ECMWF's subseasonal hindcasts beyond week 4 by improving Scandinavian Blocking (SB) and Atlantic Ridge (AR) predictions. Analysis of high-confidence predictions reveals that NAO-, SB, and AR opportunity forecasts can be associated with SPV variability and MJO phase patterns aligning with established pathways, also indicating new patterns. Overall, our work demonstrates that encoding physically meaningful climate fields can enhance S2S prediction skill, advancing AI-driven subseasonal forecast. Moreover, the experiments highlight the potential of deep learning methods as investigative tools, providing new insights into atmospheric dynamics and predictability.

Figures (9)

• Figure 1: Maps of the NAE regimes computed for ERA$5$ data between $1980$ to $2023$ (see hannachi2017 for comparison). The regime name and percentage of occurrence within the data are provided in the panel titles.
• Figure 2: Schematic of the three DL architectures, arranged vertically by increasing complexity: A) LSTM, B) Index-LSTM, and C) ViT-LSTM. Each model is designed to predict the probabilities of North Atlantic-European (NAE) regimes for the next six weeks, given NAE regime sequences from the previous six weeks (input: left panel; output: right panel). LSTM only uses NAE regime classifications as input. Index-LSTM extends this approach by incorporating additional dynamical predictors: the Madden-Julian Oscillation (MJO) phase index as a one-hot encoded vector and the stratospheric polar vortex (SPV) index as a continuous variable (additional input panel in B). ViT-LSTM represents the most advanced architecture, replacing predefined teleconnection indices with spatial climate fields. Instead of receiving MJO and SPV indices, it directly processes zonal wind (u10) in the polar region and the outgoing longwave radiation (olr) in the tropics, both for the past six weeks. These fields are encoded using a Vision Transformer (ViT) encoder, which extracts spatial features. These are then combined with the regime class information and passed to the LSTM-decoder, enabling the model to learn from spatial information potentially not captured by the conventional indices that influence S2S regime variability.
• Figure 3: Analysis of class-wise mean forecast skill across regimes (different subplots with combined performance in the first plot) and predicted lead weeks (x-axis) using A) Accuracy and B) CSI (Critical Success Index)
• Figure 4: Analysis of how past NAE regimes influence model prediction probability of the regimes at different lead weeks. In each panel, we can observe the relative prediction probability of regime B (subplot titles) at a specific lead week ($A,B \in \{\text{SB}, \text{NAO}-,\text{AR},\text{NAO}+\}$), given a regime A (y-axis) occurred several weeks before the regime B was predicted (Lag on x-axis). The values represent the prediction probability of regime B relative to the average regime frequency of regime B, highlighting precursor regime occurrence patterns beyond the expected frequency. The x-axis indicates the number of weeks before regime prediction, while rows within each subplot reflect a one-week lag shift in regime history based on their increased lead week prediction. To enhance interpretability, input, and output weeks were expanded to eleven time steps, allowing a clearer view of long-term dependencies. Panels A, B, and C correspond to LSTM, Index-LSTM, and ViT-LSTM, respectively.
• Figure 5: SPV index anomalies (relative to the mean spv index per class) over the lag in weeks for each predicted NAE regime (columns) for A) LSTM, B)Index-LSTM, C)ViT-LSTM. The black dashed lines show the mean SPV index anomaly evolution across the six input weeks (different markers) for high-probability predicted regimes. The violet line represents the SPV anomalies across all regime predictions, that is true and false positives. The green line shows the SPV anomalies across all target events in the test set, coinciding for all models. Shading of the lines denotes standard deviation (across $100$ models). Strong vortex states, defined as exceeding the 80th percentile are indicated by the blue background shading, while weak SPV states, below the 30th percentile, are indicated in red Tripathi2015.