In this paper, we present Pangu-Weather, a deep learning based system for fast and accurate global weather forecast. For this purpose, we establish a data-driven environment by downloading $43$ years of hourly global weather data from the 5th generation of ECMWF reanalysis (ERA5) data and train a few deep neural networks with about $256$ million parameters in total. The spatial resolution of forecast is $0.25^\circ\times0.25^\circ$, comparable to the ECMWF Integrated Forecast Systems (IFS). More importantly, for the first time, an AI-based method outperforms state-of-the-art numerical weather prediction (NWP) methods in terms of accuracy (latitude-weighted RMSE and ACC) of all factors (e.g., geopotential, specific humidity, wind speed, temperature, etc.) and in all time ranges (from one hour to one week). There are two key strategies to improve the prediction accuracy: (i) designing a 3D Earth Specific Transformer (3DEST) architecture that formulates the height (pressure level) information into cubic data, and (ii) applying a hierarchical temporal aggregation algorithm to alleviate cumulative forecast errors. In deterministic forecast, Pangu-Weather shows great advantages for short to medium-range forecast (i.e., forecast time ranges from one hour to one week). Pangu-Weather supports a wide range of downstream forecast scenarios, including extreme weather forecast (e.g., tropical cyclone tracking) and large-member ensemble forecast in real-time. Pangu-Weather not only ends the debate on whether AI-based methods can surpass conventional NWP methods, but also reveals novel directions for improving deep learning weather forecast systems.
核心贡献 · Key contributions
首个在所有天气要素和 1 小时至 1 周预报时间上超越业务 IFS 的 AI 方法。 First AI-based method to surpass operational IFS in all weather factors and forecast times from 1 hour to 1 week.
提出 3D 地球专用 Transformer（3DEST），将高度信息整合为 3D 数据以改进大气建模。 Proposes 3D Earth Specific Transformer (3DEST) integrating height information into 3D data for better atmospheric modeling.
引入分层时间聚合，通过训练多个不同预报时效的模型来减少累积预报误差。 Introduces hierarchical temporal aggregation to reduce cumulative forecast errors by training models with multiple lead times.
实现 0.25°空间分辨率，单 GPU 推理速度比业务 IFS 快 10000 倍以上。 Achieves 0.25° spatial resolution with inference speed over 10,000x faster than operational IFS on a single GPU.
在极端天气预报和大集合预报场景中展现出强劲性能。 Demonstrates strong performance in extreme weather forecast and large-member ensemble forecast scenarios.
终结了 AI 方法能否在全球天气预报中超越传统 NWP 方法的争论。 Ends the debate on whether AI-based methods can surpass conventional NWP methods for global weather forecast.
局限 · Limitations
训练数据仅限 39 年 ERA5 再分析数据；更长时段可能提升精度。 Training data limited to 39 years of ERA5 reanalysis; longer periods may improve accuracy.
受算力限制，模型在 100 个 epoch 后未完全收敛。 Model not fully converged after 100 epochs due to computational budget constraints.
3DEST 架构需要大量 GPU 内存，限制了网络深度和输入特征。 3DEST architecture requires large GPU memory, limiting network depth and input features.
分层时间聚合使用固定预报时效；自适应选择可能进一步减少误差。 Hierarchical temporal aggregation uses fixed lead times; adaptive selection may further reduce errors.
极端天气预报倾向于低估强度，尤其在较长预报时效上。 Extreme weather forecast tends to underestimate intensity, especially for longer forecast ranges.
论文章节 · Sections（共 16）
摘要Abstract
1 引言1 Introduction
2.1 问题设定与符号2.1 Problem Setting and Notations
2.2 NWP 方法2.2 NWP Methods
2.3 基于 AI 的方法2.3 AI-based Methods
2.4 见解2.4 Insights
3.1 概述3.1 Overview
3.2 数据准备与预训练任务3.2 Data Preparation and the Pre-training Task
3.3 3D 地球专用 Transformer3.3 3D Earth-Specific Transformer