Extreme temperature events are becoming more frequent and intense as climate change accelerates. Yet predicting these abrupt, non-stationary phenomena remains a fundamental challenge for conventional time series models, which often fail to capture rapid transitions and anomalous patterns that deviate significantly from historical behaviors.
Now, researchers from the Hangzhou Institute for Advanced Study at the University of Chinese Academy of Sciences have developed Hankelformer, a novel deep learning architecture that dramatically improves the forecasting of extreme weather events. The findings are published in National Science Review.
Hankelformer introduces two key innovations. First, it employs a structured augmentation module that constructs Hankel matrices to capture local spatiotemporal dynamics without disrupting temporal coherence, generating delay-embedding-inspired views that are topologically equivalent to the original input sequences. Second, it uses a dual-stream contrastive learning framework in which both the original and augmented sequences are processed through shared-weight Transformer encoders, maximizing agreement between the two representations. This approach significantly enhances feature invariance and robustness against distribution shifts.
The team evaluated Hankelformer on nine benchmark datasets spanning energy, transportation, and extreme weather domains. On three custom datasets capturing real-world extreme events—the 2021 Texas winter storm (TexasFreeze), the 2021 Pacific Northwest heat dome (Heatwave), and the 2020 Antarctic Peninsula heat event (Antarctic Heat)—Hankelformer consistently achieved state-of-the-art performance. Compared to leading baselines, the model delivered up to 34% improvement in Mean Squared Error (MSE).
In validation tests on a 90-dimensional chaotic Lorenz system, Hankelformer demonstrated exceptional noise robustness, maintaining low prediction error even under strong Gaussian noise interference. Ablation studies confirmed that both the Hankel augmentation and the contrastive learning components are essential: using augmentation without the contrastive loss actually degraded performance, highlighting the critical role of contrastive learning in aligning heterogeneous representations and stabilizing optimization.
The success of Hankelformer provides a promising tool for reliable extreme temperature forecasting and underscores the value of constructing topologically equivalent sequences for spatiotemporal representation learning in handling real-world non-stationary time series. Beyond climate monitoring, the framework shows potential for applications in energy system management, traffic flow prediction, and other safety-critical domains where prediction failures can have severe consequences.
