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

TL;DR

This work tackles long-term tropical cyclone intensity forecasting by reframing the problem in a discrete latent space conditioned on ERA5 data and enhanced by physical constraints. The two-stage VQLTI framework first maps TC intensity to discrete latent variables via a CVQ-VAE, then performs latent-space forecasting stabilized by PI-based physical bounds and FengWu forecast fields before decoding to intensity. Key contributions include (i) discrete latent representation to better align intensity with spatial patterns, (ii) integration of a machine-learning weather prediction forecast (FengWu) as physical context, and (iii) PI-based constraints to reduce error accumulation, yielding state-of-the-art 24–120 h forecasts with substantial MSW improvements globally and in the Western North Pacific, along with robust real-time performance. The approach offers a low-inference-cost, end-to-end pipeline that can enhance operational TC warning systems by delivering accurate long-range forecasts with reduced reliance on post-processing.

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.

Figures (4)

• Figure 1: Past methods, such as LSTM, would map the same TC intensity to the same position in the latent space, thereby ignoring their spatial information. To preserve this spatial information, we use ERA5 data as a condition to map the TC intensity to a discrete latent space, such that the same TC intensity is distributed with a certain intra-class distance, thereby capturing their spatial differences.
• Figure 2: For typhoon Hinnamnor at 2022-09-02 18:00, the ERA5 data and FengWu forecast data for 2-meter temperature are visualized, along with their differences. The figure shows FengWu's 6-hour and 120-hour forecasts, with starting times of 2022-09-02 12:00 and 2022-08-28 18:00:00 respectively. As the forecast lead time increases, the tracked TC position from FengWu deviates more, resulting in spatial misalignment between the forecast and actual ERA5 data, as indicated by the red boxes.
• Figure 3: VQLTI Framework. (a) The first stage involves CVQVAE pre-training, which uses ERA5 data as the condition to map TC intensity to a discrete latent space. (b) The second stage is the forecasting stage. Based on the pre-trained model, the blue module is frozen, and the red module is unfrozen. The TC intensity information sequence and the embedding vector in the codebook are matched to obtain the discrete latent variable sequence. When predicting the future latent variables, the Physical Information (PI) is calculated based on the FengWu forecast results to physically constrain the latent variables. Finally, the forecast latent variables are decoded into TC intensity using the FengWu forecast field as the condition.
• Figure 4: T-SNE visualization of latent variables. The same color represents the same intensity value, and here we use MSW (knot) as an example.