Data Assimilation using ERA5, ASOS, and the U-STN model for Weather Forecasting over the UK

Wenqi Wang, Jacob Bieker, Rossella Arcucci, César Quilodrán CasasarXiv 2024 · DOI · arXiv

Abstract

This paper explores the integration of Machine Learning (ML) with Data Assimilation (DA) to improve regional weather forecasting in the UK. By combining a U-Net-based model with a Sigma-Point Ensemble Kalman Filter (SPEnKF), the researchers demonstrate that while global models can be successfully adapted to local climates, the direct assimilation of sparse ground-level observations into atmospheric models remains a significant challenge that can sometimes degrade forecast accuracy.

Introduction

Traditional Numerical Weather Prediction (NWP) models are the gold standard for forecasting, but they are computationally expensive and difficult to update with real-time observations without restarting the entire simulation. Machine Learning models offer a faster, data-driven alternative, but they often struggle to incorporate new, "off-grid" observations once trained.

The authors tackle the "regionalization" problem: how to take a global ML weather model and tune it for a specific area (the UK) while effectively using Data Assimilation to merge real-time ground observations (ASOS) with atmospheric reanalysis data (ERA5). The core challenge lies in the mismatch between dense, gridded atmospheric data and sparse, irregularly distributed surface stations.

Datasets and Spatial Distribution

The study utilized two primary data sources: the ECMWF’s high-resolution ERA5 reanalysis dataset (providing hourly T850 and 2-meter temperature data) and the Automated Surface Observing System (ASOS), a network of 112 ground-based stations across the UK. Additionally, synthetic "virtual" observations were generated using the model itself to test the DA process under controlled conditions.

Comparison of sparse ASOS ground station locations (a) and dense ERA5 gridded temperature data (b) over the UK.

Methods

The researchers employed a hybrid approach combining deep learning and classical state estimation:

  • The Model (U-STN12): A U-Net architecture enhanced with a "deep spatial transformer" (U-STN), retrained for the UK region to predict 850 hPa temperature (T850) autoregressively with a 12-hour time step.
  • Data Assimilation (SPEnKF): A Sigma-Point Ensemble Kalman Filter was used to integrate observations, employing an "unscented transformation" to handle the nonlinearities inherent in weather systems.
  • Spatial Mapping: Kriging interpolation with a polynomial drift term bridged the gap between sparse ground stations and the model's grid, accounting for geographical trends in latitude and longitude.

Key Results and Insights

The U-STN12 model showed a consistent reduction in RMSE between 12-hour and 48-hour lead times, proving its effectiveness as a regional forecaster. Introducing Data Assimilation at the 24-hour mark resulted in a visible drop in error when using atmospheric data. However, a major discovery was that directly assimilating surface temperature (T2m) into an atmospheric model (T850) actually impaired performance, causing RMSE spikes at the point of assimilation.

Key insights from the experiments include:

  • Synthetic Data Utility: Using model-generated synthetic observations led to better performance than Gaussian noise, suggesting DA is most effective when data respects the learned physical manifold.
  • Surface-Atmosphere Mismatch: The difficulty of mapping sparse, non-uniformly distributed ground stations onto a dense grid can introduce errors that outweigh the benefits of the observations.
  • Regional Adaptability: Global ML architectures can be successfully "downscaled" to capture local climate nuances that global-only models might miss.

Limitations and Future Work

The authors identify several hurdles for ML-based Data Assimilation. The current 112 ASOS stations are too sparse for high-resolution requirements. Future work aims to incorporate denser datasets and move toward multi-layer models, such as Graph Neural Networks (GNNs), to better bridge the relationship between surface observations and upper-atmosphere dynamics. Additionally, further hyperparameter tuning and the development of self-correcting feedback loops are needed to minimize loss accumulation in long-term forecasts.

Summary updated

Abstract

In recent years, the convergence of data-driven machine learning models with Data Assimilation (DA) offers a promising avenue for enhancing weather forecasting. This study delves into this emerging trend, presenting our methodologies and outcomes. We harnessed the UK's local ERA5 850 hPa temperature data and refined the U-STN12 global weather forecasting model, tailoring its predictions to the UK's climate nuances. From the ASOS network, we sourced T2m data, representing ground observations across the UK. We employed the advanced kriging method with a polynomial drift term for consistent spatial resolution. Furthermore, Gaussian noise was superimposed on the ERA5 T850 data, setting the stage for ensuing multi-time step synthetic observations. Probing into the assimilation impacts, the ASOS T2m data was integrated with the ERA5 T850 dataset. Our insights reveal that while global forecast models can adapt to specific regions, incorporating atmospheric data in DA significantly bolsters model accuracy. Conversely, the direct assimilation of surface temperature data tends to mitigate this enhancement, tempering the model's predictive prowess.
BibTeX
@misc{wang2024data,
  author = {Wenqi Wang and Jacob Bieker and Rossella Arcucci and César Quilodrán Casas},
  title = {{Data Assimilation using ERA5, ASOS, and the U-STN model for Weather Forecasting over the UK}},
  year = {2024},
  eprint = {2401.07604},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2401.07604},
}
Figure