人工智能气象预报发展范式评价与融合路径

Development Paradigms Evaluation and Integration Pathways of AI Weather Forecasting

摘要:

随着再分析资料、深度学习方法和高性能计算技术的发展,人工智能（Artificial intelligence,AI）气象预报已由预报结果订正拓展至全球确定性与集合预报,以及观测资料进入预报前端的学习型预报系统。然而,现有研究多按照网络结构、预报尺度或特定任务对AI气象预报归纳,在方法分类与性能评价上仍缺少统一分析框架,难以全面反映不同技术路线的适用条件与应用范围。AI预报模型已具备全球多变量、多时效预报和快速推理能力,但在数据与状态表征、预报生成统计平滑与物理约束、性能评价与业务验证衔接以及运行适配与版本管理等方面仍存在不足。针对上述问题,本文梳理AI气象预报的技术演进,提出一种以实际预报输入为主要依据、以AI承担的核心计算环节为辅助判据分类框架,将现有方法归纳为预报产品驱动的后处理与融合方法、分析场初始化状态预报模型和观测前端驱动的学习型预报系统3类发展范式;从总体预报技巧、极端事件、空间过程、物理一致性和概率可靠性等方面构建多维性能评价框架,并结合典型研究分析评价证据成立条件及其应用情况。在此基础上,进一步诊断数据、模型、评价与业务运行等环节的主要瓶颈,提出多源、跨尺度与多圈层资料协同,训练目标改进与分层物理耦合,条件一致多维检验,以及持续监测、模型适配与版本管理相衔接的融合路径,为AI气象预报的能力评价、方法改进和业务应用提供参考。

Abstract:

Driven by advancements in reanalysis data, deep learning, and high-performance computing, AI-based weather forecasting has evolved from simple post-processing corrections to global deterministic and ensemble forecasting, and further to learning-based systems that ingest observational data directly. However, existing studies typically categorized these methods based on network architectures, forecast scales, or specific tasks. The absence of a unified framework for classification and performance evaluation hindered a comprehensive understanding of the applicability of different technical approaches. Although current AI models demonstrated capabilities in global multivariate, multi-lead-time forecasting and rapid inference, they faced challenges regarding data and state representation, statistical smoothing, physical constraints, and the alignment between performance evaluation and operational verification. To address these issues, the technological evolution of AI weather forecasting was reviewed and a classification framework was proposed based primarily on forecasting inputs and secondarily on the core computational roles of AI. The existing methods were categorized into three paradigms: forecast product-driven post-processing and fusion methods, analysis-initialized state forecasting models, and observation-driven learning-based forecasting systems. Furthermore, a multidimensional performance evaluation framework covering overall skill, extreme events, spatial processes, physical consistency, and probabilistic reliability was constructed. By analyzing typical studies, the capability boundaries and application contexts of these methods were defined. Finally, key bottlenecks in data, modeling, evaluation, and operations were identified, proposing a convergence pathway that integrated multi-source data synergy, hierarchical physical coupling, consistent operational verification, and robust version management, thereby providing a reference for the advancement and operational deployment of AI weather forecasting.