VQLTI: Long-Term Tropical Cyclone Intensity Forecasting with Physical Constraints

Xinyu Wang, Lei Liu, Kang Chen, Tao Han, Bin Li, Lei Bai

Abstract

Tropical cyclone (TC) intensity forecasting is crucial for early disaster warning and emergency decision-making. Numerous researchers have explored deep-learning methods to address computational and post-processing issues in operational forecasting. Regrettably, they exhibit subpar long-term forecasting capabilities. We use two strategies to enhance long-term forecasting. (1) By enhancing the matching between TC intensity and spatial information, we can improve long-term forecasting performance. (2) Incorporating physical knowledge and physical constraints can help mitigate the accumulation of forecasting errors. To achieve the above strategies, we propose the VQLTI framework. VQLTI transfers the TC intensity information to a discrete latent space while retaining the spatial information differences, using large-scale spatial meteorological data as conditions. Furthermore, we leverage the forecast from the weather prediction model FengWu to provide additional physical knowledge for VQLTI. Additionally, we calculate the potential intensity (PI) to impose physical constraints on the latent variables. In the global long-term TC intensity forecasting, VQLTI achieves state-of-the-art results for the 24h to 120h, with the MSW (Maximum Sustained Wind) forecast error reduced by 35.65%-42.51% compared to ECMWF-IFS.

Introduction

Predicting the intensity of tropical cyclones (TCs)—measured by Maximum Sustained Wind (MSW) and Minimum Sea Level Pressure (MSLP)—is vital for disaster preparedness. While Numerical Weather Prediction (NWP) models are the traditional gold standard, they are computationally expensive. Conversely, existing deep learning (DL) models are faster but often struggle with long-term (24h to 120h) accuracy.

The VQLTI framework addresses two main gaps in current DL approaches:

  1. Spatial Ignorance: Many models map identical TC intensities to the same point in a latent space, ignoring the environmental context.
  2. Lack of Physical Grounding: Purely data-driven models often ignore the laws of physics, leading to "error accumulation" where small mistakes early in a forecast spiral into unrealistic predictions days later.

Methods

VQLTI introduces a two-stage architecture designed to bridge the gap between data-driven flexibility and physical reality.

Stage 1: Pre-training with Vector Quantization (VQ)

The model uses a Conditional Vector Quantized-Variational AutoEncoder (CVQ-VAE). Instead of a continuous latent space, it maps TC intensity into a discrete "codebook." Crucially, this mapping is conditioned on high-resolution spatial data (ERA5). This ensures that the latent representation of a storm captures both its intensity and its unique spatial environment.

Stage 2: Forecasting with Physical Constraints

To predict the future, VQLTI employs a Latent Variable Iteration Model. It incorporates "future" environmental conditions provided by FengWu, a state-of-the-art DL weather prediction model. To prevent the model from drifting into physically impossible territory, the authors calculate the Potential Intensity (PI)—a theoretical thermodynamic limit on how strong a storm can get based on sea surface temperature and atmospheric conditions. This PI acts as a "physical guardrail" for the latent variables during the iterative forecasting process.

The two-phase VQLTI framework: (a) Pre-training the discrete latent space and (b) the iterative forecasting stage incorporating physical constraints.

Key Results and Insights

VQLTI achieves state-of-the-art (SOTA) performance across global basins for 24h to 120h leads. Compared to the ECMWF-IFS (a top-tier operational model), VQLTI reduced MSW forecast errors by 35.65% to 42.51%. It outperformed recent deep learning benchmarks like TC-Pre and MSCAR, particularly in long-term scenarios where those models typically suffer from rapid error accumulation.

  • Discrete Latent Spaces for Diversity: By using a discrete codebook (VQ), the model can represent the same intensity value in multiple ways depending on the surrounding weather patterns. This "intra-class distance" allows the model to be much more sensitive to spatial nuances.
  • Physics as a Ceiling: The inclusion of Potential Intensity (PI) prevents the model from overestimating intensity in environments that physically cannot support a stronger storm.
  • Latent Space Clarity: T-SNE visualizations confirm that the VQ-quantized latent space successfully clusters storms by intensity while maintaining the necessary spread to account for different spatial environments.
T-SNE visualization comparing latent spaces with and without Vector Quantization (VQ).

Conclusion

VQLTI demonstrates that combining discrete latent representations with physical constraints significantly improves long-term TC intensity forecasting. While the model currently depends on external forecast fields like FengWu and focuses primarily on intensity rather than track, it sets a new benchmark for deep-learning-based meteorological forecasting. Future work will look into tighter integration with TC track forecasting and optimizing the computational requirements of the two-stage training process.

Summary updated

Abstract

Tropical cyclone (TC) intensity forecasting is crucial for early disaster warning and emergency decision-making. Numerous researchers have explored deep-learning methods to address computational and post-processing issues in operational forecasting. Regrettably, they exhibit subpar long-term forecasting capabilities. We use two strategies to enhance long-term forecasting. (1) By enhancing the matching between TC intensity and spatial information, we can improve long-term forecasting performance. (2) Incorporating physical knowledge and physical constraints can help mitigate the accumulation of forecasting errors. To achieve the above strategies, we propose the VQLTI framework. VQLTI transfers the TC intensity information to a discrete latent space while retaining the spatial information differences, using large-scale spatial meteorological data as conditions. Furthermore, we leverage the forecast from the weather prediction model FengWu to provide additional physical knowledge for VQLTI. Additionally, we calculate the potential intensity (PI) to impose physical constraints on the latent variables. In the global long-term TC intensity forecasting, VQLTI achieves state-of-the-art results for the 24h to 120h, with the MSW (Maximum Sustained Wind) forecast error reduced by 35.65%-42.51% compared to ECMWF-IFS.
BibTeX
@inproceedings{wang2025vqlti,
  author = {Xinyu Wang and Lei Liu and Kang Chen and Tao Han and Bin Li and Lei Bai},
  title = {{VQLTI: Long-Term Tropical Cyclone Intensity Forecasting with Physical Constraints}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=J2nyiCPjJc},
}
Figure