Core Concepts
Data-driven ensemble models improve sub-seasonal forecasting accuracy.
Abstract
Abstract:
Presents an operations-ready multi-model ensemble weather forecasting system.
Outperforms ECMWF extended-range ensemble for 2-meter temperature predictions.
Demonstrates near-state-of-the-art subseasonal-to-seasonal forecasts using data-driven models.
Introduction:
Early toy models evolved into accurate data-driven weather prediction systems.
Limitations include lack of testing on extended-range forecasting and focus on deterministic forecasts.
Methodology:
Trained on ECMWF ERA5 reanalysis dataset with various weather parameters.
Utilizes five trained models based on different architectures for predictions.
Results:
Compares performance with ECMWF ensemble, showing competitive results.
Bias correction improves ECMWF results more than the data-driven ensemble.
Combination of data-driven and NWP forecasts leads to better probabilistic forecasts.
Discussion:
Data-driven models show promise but face challenges in extreme event forecasting.
Multi-model approach leverages different model strengths for improved forecasts.
Stats
We present an operations-ready multi-model ensemble weather forecasting system which uses hybrid data-driven weather prediction models coupled with the European Centre for Medium-range Weather Forecasts (ECMWF) ocean model to predict global weather at 1-degree resolution for 4 weeks of lead time. For predictions of 2-meter temperature, our ensemble on average outperforms the raw ECMWF extended-range ensemble by 4-17%, depending on the lead time. However, after applying statistical bias corrections, the ECMWF ensemble is about 3% better at 4 weeks. For other surface parameters, our ensemble is also within a few percentage points of ECMWF’s ensemble. We demonstrate that it is possible to achieve near-state-of-the-art subseasonal-to-seasonal forecasts using a multi-model ensembling approach with data-driven weather prediction models.
Many of the models have not been tested on extended-range forecasting in the two- to six-week subseasonal-to-seasonal (S2S) time scales. Despite the importance of these time scales for a variety of applications such as agriculture and risk management, the models cannot be expected to generalize. While some data-driven models, including the DLWP ensemble of [Weyn et al., 2021], FuXi-S2S [Chen et al., 2023c], the SFNO [Bonev et al., 2023], and NeuralGCM [Kochkov et al., 2024], have been run at S2S or longer time scales, some key considerations, including the need for post-processing, that have not been fully addressed.
The use of regression losses in training data-driven weather models has resulted in deterministic forecasts which overly smooth fine-scale features, limiting the models’ ability to capture the full range of weather phenomena. It also “games” the target metrics such as root-mean-squared-error (RMSE), which are optimized by ensemble-mean-like forecasts, making it difficult to benchmark the actual skill of different models (some models perform particularly well on RMSE at longer lead times because of this smoothing).
Our goal here is to broadly demonstrate the performance of our ensemble of data-driven weather prediction models. To compare with the latest 100-member ECMWF IFS extended-range ensemble, we limit our test set to forecasts issued on Mondays and Thursdays at 00 UTC time starting on July 3, 2023 until Dec 31, 2023. All verification data are from...
Using a multi-model ensembling approach, we have shown that near-state-of-the-art sub-seasonal-to-seasonal forecasts are possible using data-driven weather forecasting models. Our ensemble is competitive with...
Our results show promise for operationalizing ensembles of data-driven weather forecasting models. However,...
Quotes
"We present an operations-ready multi-model ensemble weather forecasting system."
"Our goal here is to broadly demonstrate the performance..."
"Using a multi-model ensembling approach..."