TL;DR
This paper introduces VQLTI, a novel deep learning framework for long-term tropical cyclone (TC) intensity forecasting. By mapping TC intensity to a discrete latent space and incorporating physical constraints derived from the FengWu weather model, VQLTI significantly improves forecast accuracy and reduces error accumulation compared to existing methods.
What Problem Does This Paper Address?
Accurate forecasting of tropical cyclone (TC) intensity is critical for disaster preparedness and emergency response. While deep learning (DL) methods offer computational advantages over traditional Numerical Weather Prediction (NWP) models, they often struggle with long-term forecasts, exhibiting significant error accumulation. Existing DL approaches also face challenges in effectively integrating spatial information with TC intensity and incorporating crucial physical knowledge. This paper aims to bridge this gap by developing a DL model that enhances long-term forecasting accuracy through better spatial-intensity matching and the integration of physical constraints.
Methods
The VQLTI framework is built upon two key strategies:
- Discrete Latent Space Representation: VQLTI maps TC intensity information into a discrete latent space using a Conditional Vector Quantized-Variational AutoEncoder (CVQ-VAE) framework. This approach, inspired by data discretization techniques, aims to better align TC intensity with its spatial characteristics, addressing the issue where similar intensities can have diverse spatial patterns. The model uses large-scale spatial meteorological data (ERA5) as a condition for this mapping.
- Physical Knowledge Integration:
- FengWu Forecast Leverage: The framework incorporates forecasts from the FengWu machine learning-based weather prediction model. This provides additional physical knowledge that complements the ERA5 data.
- Potential Intensity (PI) Constraints: The Potential Intensity (PI), a theoretical upper limit of TC intensity under given environmental conditions, is calculated using the FengWu forecast field. This PI value is then used to impose physical constraints on the latent variables during forecasting, helping to mitigate the underestimation of TC intensity often observed in forecast fields.
The VQLTI model training involves two stages: pre-training (CVQ-VAE) and forecasting. In the forecasting stage, the model learns to predict future latent variables and then decodes them into TC intensity forecasts, leveraging the FengWu forecast field as a condition. The framework allows for forecasts at arbitrary time steps by adjusting the iteration count in the latent space.
Datasets
- Ground Truth: IBTRACS dataset (Knapp et al., 2010) for TC intensity (Maximum Sustained Wind - MSW, and Minimum Sea Level Pressure - MSLP).
- Spatial Meteorological Data: ERA5 reanalysis data (Hersbach et al., 2020) for 69 atmospheric variables.
- Physical Knowledge Source: FengWu weather prediction model forecasts.
The paper uses data from 1980-2017 for training, 2018 for validation, and 2019-2020 for testing. For real-time forecasting experiments, FengWu forecasts for 2021-2022 are generated.
Key Insights and Ideas
- Discretization for Spatial Alignment: Mapping TC intensity to a discrete latent space, rather than a continuous one, is crucial for capturing the nuances of spatial information associated with different intensities. This helps prevent the model from treating identical intensity values with vastly different spatial configurations as the same.
- Physics-Informed DL for Long-Term Forecasts: Integrating physical knowledge, particularly through the use of advanced ML weather models like FengWu and theoretical constraints like PI, is essential for overcoming the error accumulation problem in long-term DL forecasts.
- FengWu as a Proxy for Physical Understanding: The FengWu model's ability to provide skillful medium-range forecasts suggests it captures underlying physical dynamics. Leveraging its outputs as conditions and for deriving constraints is a novel way to imbue DL models with this physical understanding.
- PI as a Bound for Improvement: Using Potential Intensity as a physical constraint effectively addresses the tendency of forecast fields to underestimate TC intensity, leading to more robust long-term predictions.
Key Results
- State-of-the-Art Performance: VQLTI achieves state-of-the-art (SOTA) results in global long-term TC intensity forecasting for lead times of 24h to 120h.
- Reduced Error Accumulation: Compared to other DL methods like TC-Pre, MSCAR, and TCIF, VQLTI exhibits significantly lower error accumulation rates. For MSLP, its 48h error increases by only 16.01% compared to the 24h error, much lower than the 63.16%-69.77% seen in other methods.
- NWP Model Improvement: VQLTI reduces the 24h to 120h MSW forecast error by 35.65%-42.51% compared to ECMWF-IFS and by 25.69%-39.38% compared to NCEP-GFS.
- Real-time Forecasting: The VQLTI (real-time) model, using available data, still outperforms other DL methods and FengWu itself, demonstrating practical applicability.
- Ablation Study Validation: Ablation studies confirm the significant contributions of VQ (discretization), FengWu data, and PI constraints to the model's performance.
Limitations and Future Work
- FengWu Forecast Discrepancies: The paper acknowledges that FengWu's forecasts have inherent discrepancies compared to ERA5, which can lead to deviations in TC positions and spatial information misalignment, impacting real-time forecast accuracy.
- Future Work: The authors suggest that further improvements could be made by exploring more advanced methods for integrating physical knowledge and potentially by using higher-resolution forecast data. The framework's flexibility in handling arbitrary time steps without retraining is highlighted as a key advantage for future applications.
Summary updated