基于改进扩散模型的温度预报  

Temperature prediction based on improved diffusion model

作　　者：方巍[1,2,3,4] 袁众 薛琼莹 FANG Wei;YUAN Zhong;XUE Qiongying(School of Computer Science,Nanjing University of Information Science&Technology,Nanjing 210044,China;Engineering Research Center of Digital Forensics(Nanjing University of Information Science&Technology),Ministry of Education,Nanjing 210044,China;Collaborative Innovation Center of Atmospheric Environment and Equipment Technology(Nanjing University of Information Science&Technology),Nanjing 210044,China;Jiangsu Provincial Key Laboratory for Computer Information Processing Technology(Soochow University),Suzhou,Jiangsu 215000,China)

机构地区：[1]南京信息工程大学计算机学院,南京210044 [2]数字取证教育部工程研究中心(南京信息工程大学),南京210044 [3]大气环境与装备技术协同创新中心(南京信息工程大学),南京210044 [4]江苏省计算机信息处理技术重点实验室(苏州大学),江苏苏州215000

出　　处：《中国科技论文》2024年第2期215-223,共9页China Sciencepaper

基　　金：国家自然科学基金资助项目(42075007);江苏省计算机信息处理技术重点实验室开放课题资助项目(KJS2275)。

摘　　要：针对传统数值预报模式计算时间长和计算资源消耗大的问题,以及现有深度学习预报方法在温度预报结果上不精确,且预测结果模糊的问题,提出了一个新的温度预报模型。首先,设计了一个时空信息捕捉模块,将该模块捕获的长期依赖信息,作为扩散模型的生成条件,赋予扩散模型预报的能力;其次,设计了一个新的平衡损失函数,同时保护了扩散模型的生成能力和时空信息捕捉模块对时空信息的捕捉能力;最后,基于美国国家环境预报中心的再分析数据进行预报,与现有的深度学习方法相比,所提模型预报结果的质量在均方误差(mean square error,MSE)上降低了17.3%,在均方根误差(root mean square error,RMSE)上降低了9.14%,在峰值信噪比(peak signal to noise ratio,PSNR)上提升了5.1%。改进的扩散模型能有效地捕捉时空依赖的关系,有效地进行时空序列预测,效果优于其他对比方法。Traditional numerical forecasting models suffer from long computation time and high resource consumption,while the existed deep learning forecasting methods for temperature prediction have the disadvantages of inaccuracies and dissipation.To address these problems,a new architecture was proposed.First,a spatiotemporal perception module was designed,and the long-term dependency information captured by the module was taken as the generation condition for the diffusion model,which endowed the diffusion model with the ability to forecast.Second,a new equilibrium loss function was designed,which protected both the gen-eration ability of the diffusion model and the spatio-temporal information capture ability of the spatiotemporal perception module.Finally,the reanalysis data from the National Centers for Environmental Prediction(NCEP)was used to perform forecasting,and the results were compared with existed deep learning methods.The quality of predictions in this paper is 17.3%lower in mean square error(MSE),9.14%lower in root mean square error(RMSE),and 5.1%higher in peak signal to noise ratio(PSNR).The improved diffusion model can effectively capture the spatio-temporal dependency relationship and perform spatio-temporal sequence prediction,which is better than other comparison methods.

关 键 词：时空序列预测 深度学习 扩散模型 时空捕捉模块 平衡损失函数 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]