Pangu-Weather and GraphCast beat official track forecasts in 2024

An ESCAP/WMO Typhoon Committee expert team verified 19 AI forecast setups for the 2024 season. The best AI models had smaller track errors than official forecasts and physics-based models.

  • 2 an organization

    Across the 2024 season, the best AI models had smaller mean position errors than official forecasts and physics-based models, more so at longer lead times.

Depth runs from 1, a convenience or time saved, to 5, a life or a livelihood preserved.

Status
Reported benefit.
Evidence

foundstudiedreportedclaimed

Only the deployer supports the benefit.

Compared with

Official forecasts by CMA, HKO, JMA, JTWC and KMA, plus the ECMWF IFS and JMA GSM models. Measured.

The system’s part
The team compared the AI models' own track forecasts with official forecasts and physics-based models on a homogeneous sample, which isolates the models' skill. The team notes that most models were trained on ERA5 reanalysis data. The verifying agencies run the models themselves, and the makers are outside them, except CMA for FengQing.
Limits
Not yet checked.
Costs and harms
Not yet checked.
Later
Not yet checked.

Download data Cite Suggest a correction

Benefit type
Improved Task Accuracy
System
19 AI typhoon forecast setups, with Pangu-Weather (HKO), GraphCast (JMA and KMA) and FengWu (Shanghai Meteorological Service) as the best track models
Made by
Huawei (Pangu-Weather), Google DeepMind (GraphCast), Shanghai Artificial Intelligence Laboratory (FengWu), Fudan University (FuXi), ECMWF (AIFS), NVIDIA (FourCastNet) and CMA (FengQing)
Used by
CMA, HKO, JMA, KMA and the Shanghai Meteorological Service, each running the models
Where
Western North Pacific, with forecasts made in China, Hong Kong, Japan and South Korea
Stage
In routine use
Date
, when the source was published
Kind
Aggregate result
Broad benefit
Estimating the Quantifiable Characteristics of Products, Events, or Information
Last revised
31 October 2026

Sources

  1. Report, by the organization that used it.

    ESCAP/WMO Typhoon Committee, Working Group on Meteorology, Expert Team on AI Applications in Tropical Cyclone Analysis and Forecasting (ET-AITC). Annual Operating Plan 13: Promoting Technical Exchange of AI Applications in Tropical Cyclone Analysis and Forecasting, Activity Report 2025. April 2026, section 3.2.

    AI-based models consistently produced the smallest mean position errors, with the advantage becoming increasingly pronounced at longer lead times.

Details

AI system

Workflow

Five agencies ran AI models from the initial conditions of global models and tracked storm centers in the output. The team verified the forecasts against RSMC Tokyo best-track data up to 120 hours.

Follow-up

Conditions for replication

The result holds over more seasons, which the team calls for. Intensity forecasts come from other sources, since the AI models underestimate intensity.

Entry details

Sector

Weather forecasting / tropical cyclones.

Purpose

Huawei, Google DeepMind and the Shanghai Artificial Intelligence Laboratory built Pangu-Weather, GraphCast and FengWu to forecast global weather in the medium range with models trained on past weather data.

Not yet checked: model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, beneficiaries, reach, distribution and access, duration, every measurement question, every money question, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.

Cite this record

Peter H. Vartanian. (2026). Pangu-Weather and GraphCast beat official track forecasts in 2024 (AIBD-0756). Artificial Intelligence Benefits Database, edition of 31 October 2026. https://petervartanian.xyz/aibd/records/AIBD-0756.html (archived at https://doi.org/10.7910/DVN/AADSII, version 4.0)

BibTeX Data Field definitions