南京大学学报（自然科学版）2025，Vol.61Issue(3)：373-386,14.DOI:10.13232/j.cnki.jnju.2025.03.003
基于潜空间扩散模型的台风路径与强度预报模型研究
On typhoon forecasting models for track and intensity based on latent diffusion model
摘要
Abstract
Accurate forecasting of typhoon track and intensity is the basis of typhoon disaster reduction and prevention.In this study,a typhoon forecast model for track and intensity is developed based on the latent diffusion model(LDM).To this end,the LDM is improved,and the structural similarity(SSIM)loss and discrete Fourier transform(DFT)loss are introduced to improve the performance of VAE in reconstructing the details of meteorological variables.In addition,the Diffusion Transformer(DiT)with better scalability is used as the noise prediction network,and the pre-training and fine-tuning strategy is used to optimize the forecast model to further enhance the forecasting ability of the model.The forecast model is used to test the 6,12,18 and 24 h forecasts of nine typhoons in the northwest Pacific in 2023.The forecast results show that the improved LDM has a significant improvement in the typhoon track and intensity forecasting ability.The ensemble method can further enhance the stability and accuracy of the forecast model,especially in typhoon track forecasting.These results provide new insights and methods for future AI-based typhoon forecasting.关键词
台风/深度学习/扩散模型/预报Key words
typhoon/deep learning/diffusion model/forecasting分类
天文与地球科学引用本文复制引用
闵隽玉,谈哲敏..基于潜空间扩散模型的台风路径与强度预报模型研究[J].南京大学学报（自然科学版）,2025,61(3):373-386,14.基金项目
国家自然科学基金(42192555) （42192555）
