Research map / Probabilistic and scientific modeling / Molecules and proteins
VQLTI: Long-Term Tropical Cyclone Intensity Forecasting with Physical Constraints
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes VQLTI, which maps tropical cyclone intensity into a discrete latent space conditioned on ERA5 spatial fields, forecasts latent codes using FengWu forecast fields and potential-intensity physical constraints, and decodes back to intensity for long-term TC forecasts.
Problem
Long-term (24h to 120h) tropical cyclone intensity forecasting suffers from poor matching between intensity and spatial information and from error accumulation due to missing physical knowledge.
Approach
Discretize TC intensity into a VQ codebook conditioned on ERA5 spatial patches (CVQ-VAE), forecast latent codes autoregressively with cross-attention, condition decoding on FengWu forecast fields, and impose potential-intensity upper bounds on the latent forecasts.
Key results
- Global 2019-2020 test: VQLTI MSW MAE 4.30-5.05 m/s over 24h-120h, outperforming ECMWF-IFS and NCEP-GFS; MSW error reduced 35.65%-42.51% vs ECMWF-IFS per the abstract.
- Real-time 2022: VQLTI (real-time) 120h MSW MAE 5.53 vs ECMWF 8.62 and NCEP 6.45 m/s.
- Ablation on 2022 real-time: adding FengWu fields cuts 24h-120h MSW error by 55.29%-70.56%; PI constraint further reduces 120h MSLP and MSW errors by 30.75% and 24.35%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add codebook quantization (ablation) | VQLTI without quantization, FengWu, or PI · Global 2022, 120h | MSLP MAE: 25.10 -> 21.45 | Helped |
| Add FengWu forecast conditioning (ablation) | VQLTI with codebook quantization · Global 2022, 120h | MSW MAE: 16.35 -> 7.31 | Helped |
| Add PI physical constraints (ablation) | VQLTI with codebook quantization and FengWu conditioning · Global 2022, 120h | MSLP MAE: 11.35 -> 7.86 | Helped |
| Use real-time analysis inputs and FengWu forecasts instead of the standard inputs | VQLTI · Global 2022, 120h | MSW MAE: 4.60 -> 5.53 | Hurt |
| Use real-time analysis inputs and FengWu forecasts instead of the standard inputs | VQLTI · Global 2022, 24h | MSW MAE: 4.51 -> 4.47 | Helped |
| Use VQLTI instead of ECMWF | Global TC intensity forecasting · Global 2019–2020, 120h | MSW MAE: 8.15 -> 5.05 | Helped |
| Use VQLTI instead of MGTCF | WP TC intensity forecasting · WP 2019, 120h | MSW MAE: 8.24 -> 5.23 | Helped |
Surprising findings
- Forecasting in the discrete latent space with FengWu conditioning yields errors far below ERA5-based reference values, suggesting the learned latent forecast corrects underestimated intensity.
- Pure time-series or ERA5-only deep models (TC-Pre, MSCAR, TCIF) accumulate errors quickly, with 48h MSLP error rising 60%+ over 24h.
Limitations
- Real-time performance degrades at long lead times due to FengWu forecast discrepancies and TC tracker misalignment.
- PI uses t2m as a substitute for sea-surface temperature, which the authors acknowledge introduces deviation.
- Deep learning baselines were run with fixed 20-step horizons and MMSTN/MGTCF ensemble means were recomputed, so comparisons may not match original papers.
Experimental setup
Datasets: IBTrACS ERA5 FengWu forecasts
Benchmarks: IBTrACS global TC 2019-2020 WP basin 2019 global real-time TC 2022
Metrics: MAE MSLP error (hPa) MSW error (m/s) skill score
Baselines: TC-Pre MSCAR TCIF ECMWF-IFS NCEP-GFS MMSTN MGTCF JTWC FengWu ERA5
Builds on: CVQ-VAE VQ-VAE MSCAR FengWu LSTM
Closest papers by meaning
- TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical Cyclone Forecasting (AAAI 2025 · similarity 0.86)
- TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness (ICML 2025 · similarity 0.85)
- IDOL: Meeting Diverse Distribution Shifts with Prior Physics for Tropical Cyclone Multi-Task Estimation (NeurIPS 2025 · similarity 0.84)
- Long-Term Typhoon Trajectory Prediction: A Physics-Conditioned Approach Without Reanalysis Data (ICLR 2024 · similarity 0.82)
- Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling (NeurIPS 2024 · similarity 0.81)
- CyPortQA: Benchmarking Multimodal Large Language Models for Cyclone Preparedness in Port Operation (AAAI 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.