Truace
Subscribe
52/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Beijing temperature forecasts: study finds AI models beat physics models, and ML blending cut error 24%

Study title: Machine Learning‐Based Postprocessing Ensemble of AI and NWP Models for Short‐Term Air Temperature Forecasting Over Beijing

Evidence type
Benchmark or lab evaluation
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2579cite or check it

Study title: Machine Learning‐Based Postprocessing Ensemble of AI and NWP Models for Short‐Term Air Temperature Forecasting Over Beijing

Researchers evaluated 6–72 hour air temperature forecasts over Beijing from January 2022 to December 2023, comparing physics-based numerical weather prediction with the AI models GraphCast and Pangu. Postprocessing the outputs with an XGBoost ensemble cut root-mean-square error by 24.1% and raised temperature forecast accuracy by 22.0% against a simple ensemble mean.

Read the study abstract

ABSTRACT Accurate short‐term air temperature forecasts are critical for societal and economic activities, particularly for energy management and urban planning. This study evaluates 6–72 h air temperature forecasts over Beijing from January 2022 to December 2023 using both numerical weather prediction (NWP) and AI‐based models. The AI models, GraphCast and Pangu, were initialized with the ECMWF Integrated Forecasting System (IFS) and NOAA Global Forecast System (GFS), resulting in GraphCast‐IFS, Pangu‐IFS, Graph…

Machine Learning‐Based Postprocessing Ensemble of AI and NWP Models for Short‐Term Air Temperature Forecasting Over Beijing
"Cracked earth after prolonged drought. 2020" by Rosser1954, CC BY-SA 4.0.

The gain

In a study of 6–72 hour temperature forecasts over Beijing in 2022–2023, an XGBoost ensemble of AI and physics-based models cut forecast error by 24.1% and raised accuracy by 22.0% versus a simple ensemble mean.

In brief

The AI models were run from two sets of starting conditions, ECMWF's IFS and NOAA's GFS, giving four variants. The AI models consistently outperformed the physics models that drove them, especially when started from IFS, and Pangu-IFS showed the best skill of the four.

The gains were largest at stations in the mountainous west and north of Beijing. The authors say the approach may benefit operational and energy-related forecasting, where short-term temperature accuracy matters for energy management and urban planning.

Main points

  1. The study covered 6–72 hour temperature forecasts over Beijing from January 2022 to December 2023.
  2. Four AI variants were tested: GraphCast-IFS, Pangu-IFS, GraphCast-GFS and Pangu-GFS, started from ECMWF and NOAA data.
  3. AI models consistently beat the physics models that drove them, particularly when initialized with IFS.
  4. Pangu-IFS had the best predictive skill among the individual models.
  5. Three machine-learning methods were compared for postprocessing: multiple linear regression, gradient boosting regression tree and XGBoost.
  6. XGBoost cut RMSE by 24.1% and raised accuracy by 22.0% (p < 0.01) versus the simple ensemble mean.
  7. Improvements were strongest at stations in the mountainous regions of western and northern Beijing.

The full story

The work compares forecasts from physics-based models and two AI weather models, GraphCast and Pangu, each started from either ECMWF IFS or NOAA GFS data. Forecast skill varied strongly by season and time of day, but AI models consistently outperformed the physics models driving them, with the largest edge when started from IFS.

To improve results further, the researchers combined outputs from AI and physics models using a multimodel ensemble of three machine-learning algorithms: multiple linear regression, gradient boosting regression tree and XGBoost. XGBoost delivered the biggest gains, with a 24.1% lower RMSE and 22.0% higher forecast accuracy than a simple ensemble mean.

Error decomposition showed the improvements came mainly from reducing systematic bias and phase errors, with bias correction the dominant factor. The effect was most pronounced at stations in complex mountainous terrain in western and northern Beijing.

Sources

1 on file
  1. Meteorological Applicationsdoi:10.1002/met.70238
    Peer-reviewed2026-09-01

Talk it over

Four questions to start onWho gets this benefit, and who was left out of the study or story?

Start with a question

Beijing temperature forecasts: study finds AI models beat physics models, and ML blending cut error 24%https://truvace.com/item/machine-learning-based-postprocessing-ensemble-of-ai-and-nwp-models-for-short-te

  1. Who gets this benefit, and who was left out of the study or story?

    The gain In a study of 6–72 hour temperature forecasts over Beijing in 2022–2023, an XGBoost ensemble of AI and physics-based models cut forecast error by 24.1% and raised accuracy by 22.0% versus a simple ensemble mean.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on a benchmark or lab test, from Meteorological Applications.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate