Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data driven weather forecasts trained and initialised directly from observations

Anthony McNally, Christian Lessig, Peter Lean, Eulalie Boucher, Mihai Alexe, Ewan Pinnington, Matthew Chantry, Simon Lang, Chris Burrows, Marcin Chrust, Florian Pinault, Ethel Villeneuve, Niels Bormann, Sean Healy

TL;DR

It is argued that this new approach, training a neural network to predict future weather purely from historical observations with no dependence on reanalyses, avoids many of the challenges of traditional data assimilation, can exploit a wider range of observations and is readily expanded to simultaneous forecasting of the full Earth system.

Abstract

Skilful Machine Learned weather forecasts have challenged our approach to numerical weather prediction, demonstrating competitive performance compared to traditional physics-based approaches. Data-driven systems have been trained to forecast future weather by learning from long historical records of past weather such as the ECMWF ERA5. These datasets have been made freely available to the wider research community, including the commercial sector, which has been a major factor in the rapid rise of ML forecast systems and the levels of accuracy they have achieved. However, historical reanalyses used for training and real-time analyses used for initial conditions are produced by data assimilation, an optimal blending of observations with a physics-based forecast model. As such, many ML forecast systems have an implicit and unquantified dependence on the physics-based models they seek to challenge. Here we propose a new approach, training a neural network to predict future weather purely from historical observations with no dependence on reanalyses. We use raw observations to initialise a model of the atmosphere (in observation space) learned directly from the observations themselves. Forecasts of crucial weather parameters (such as surface temperature and wind) are obtained by predicting weather parameter observations (e.g. SYNOP surface data) at future times and arbitrary locations. We present preliminary results on forecasting observations 12-hours into the future. These already demonstrate successful learning of time evolutions of the physical processes captured in real observations. We argue that this new approach, by staying purely in observation space, avoids many of the challenges of traditional data assimilation, can exploit a wider range of observations and is readily expanded to simultaneous forecasting of the full Earth system (atmosphere, land, ocean and composition).

Data driven weather forecasts trained and initialised directly from observations

TL;DR

It is argued that this new approach, training a neural network to predict future weather purely from historical observations with no dependence on reanalyses, avoids many of the challenges of traditional data assimilation, can exploit a wider range of observations and is readily expanded to simultaneous forecasting of the full Earth system.

Abstract

Skilful Machine Learned weather forecasts have challenged our approach to numerical weather prediction, demonstrating competitive performance compared to traditional physics-based approaches. Data-driven systems have been trained to forecast future weather by learning from long historical records of past weather such as the ECMWF ERA5. These datasets have been made freely available to the wider research community, including the commercial sector, which has been a major factor in the rapid rise of ML forecast systems and the levels of accuracy they have achieved. However, historical reanalyses used for training and real-time analyses used for initial conditions are produced by data assimilation, an optimal blending of observations with a physics-based forecast model. As such, many ML forecast systems have an implicit and unquantified dependence on the physics-based models they seek to challenge. Here we propose a new approach, training a neural network to predict future weather purely from historical observations with no dependence on reanalyses. We use raw observations to initialise a model of the atmosphere (in observation space) learned directly from the observations themselves. Forecasts of crucial weather parameters (such as surface temperature and wind) are obtained by predicting weather parameter observations (e.g. SYNOP surface data) at future times and arbitrary locations. We present preliminary results on forecasting observations 12-hours into the future. These already demonstrate successful learning of time evolutions of the physical processes captured in real observations. We argue that this new approach, by staying purely in observation space, avoids many of the challenges of traditional data assimilation, can exploit a wider range of observations and is readily expanded to simultaneous forecasting of the full Earth system (atmosphere, land, ocean and composition).
Paper Structure (7 sections, 5 figures, 1 table)

This paper contains 7 sections, 5 figures, 1 table.

Table of Contents

  1. The current use of observations in NWP
  2. Does machine learning offer an alternative?
  3. Observation training data
  4. Prototype neural network
  5. Results of the prototype experiments
  6. Discussion and future perspectives
  7. Appendix A: Glossary of satellite sensor acronyms

Figures (5)

  • Figure 1: Typical examples of the data coverage provided by ATMS microwave radiances (a), IASI infrared radiances (b), AVHRR visible reflectances (c) and SYNOP surface measurements of 2m temperature and 10m wind (d) within a 12-hour window.
  • Figure 2: A schematic representation of the transformer neural network used in the prototype experiments. On the left we see portions of the satellite observations being randomly masked, and during training the network being challenged to predict the values of the masked data.
  • Figure 3: An example of the network predicting 12-hours of IASI window channel radiances (channel 921 around 10 microns) for a case on October 17th 2022 (00z) using the measurements from the previous 12-hour window as input. In this IASI channel, light shades correspond to cold features (e.g. clouds and the Antarctic continent) and dark shades to warm features (e.g. the satellite viewing warm surfaces unobscured by clouds).
  • Figure 4: Example of input (a), predicted (b) and target (c) METOP-B AVHRR visible reflectances in the next 12-hour window on October 17th 2022 (00z), along with the target truth. Note the accurate prediction of visible cloud features in the western hemisphere despite this hemisphere being in darkness for the input window and having no AVHRR information.
  • Figure 5: Example of 10m winds inferred from satellite brightness temperatures alone. Predicted SYNOP 10m wind (upper panels) and verifying target SYNOP 10m wind (lower panels) at 9UTC and 18UTC for a case on February 18th 2022.