The Future of Weather Forecasting

It’s difficult to make predictions, especially about the future – Danish Proverb

For many years, looking up at a red sky at night brought delight to shepherds. Cows lying in the fields, however, was a sign to prepare for rain. 

Fortunately, modern methods of weather prediction are more sophisticated. Weather prediction is critical for managing the UK’s growing renewable energy resources, understanding climate change, and improving transportation and agriculture. Generating faster and more accurate weather predictions is therefore increasingly important. 

AI is playing a big part in the field of green technology, as discussed in our recent AI report [1]. Weather prediction is no exception, with data-driven machine learning (ML) being used extensively in this space.

In this article, we provide a brief overview of the modern weather prediction method, and discuss recent research to improve weather prediction by incorporating ML techniques.

Numerical weather prediction - the “traditional” modern approach

Since the 1950s, numerical weather prediction (NWP) has been used for weather prediction. An NWP system consists of a series of models of different components of the Earth’s atmosphere. Decades of research in Earth observation, data assimilation, and fluid dynamics have resulted in the increasing accuracy and complexity of these systems. Modern NWP systems are run on purpose-built supercomputers [2].

The NWP pipeline begins with acquiring observations from various sources. The collected data is fed into a data assimilation system to be combined with an initial guess from a previous forecast to generate an approximation of the current state of the atmosphere. The approximation is then used as an initial state for a forecasting system, which integrates fluid and thermodynamic equations and outputs predictions. Finally, the predictions are used for downstream tasks, for example to generate local predictions. 

Whilst the NWP approach generates predictions of high accuracy, it is an inflexible model involving an intricate flow of information from one system to the next. Moreover, “traditional” NWP requires the enormous processing power of a supercomputer, restricting access for developing nations and small research institutions. 

Machine learning solves problems

Recent research has successfully applied ML techniques to the components of the NWP pipeline. Careful application of ML has enabled accurate predictions to be generated on commercially available hardware in a fraction of the time that it would take even the most powerful supercomputer. 

Improved performance requiring significantly less computation can be obtained by replacing the “numerical solver” – the computationally-expensive method for solving the equations in the forecasting system – with a trained ML model (such as Google DeepMind’s GraphCast). [3] 

Another interesting technique involves using ML to derive useful data for the NWP prediction from raw satellite data – for example, deriving wind direction predictions from radar images of the sea surface. [4]

These techniques look to replace or supplement components of the NWP pipeline with ML models, to great effect.

Thinking outside the pipeline

A paper published last month investigates a different approach. Known as “Aardvark Weather”, the researchers replaced the entire NWP pipeline with a single end-to-end ML model [5]. Notably, the trained model can be run on a desktop computer. The results show comparable forecasting accuracy across various meteorological parameters. 

Furthermore, the researchers state that the end-to-end ML approach enables fine-tuning of bespoke models. This improves prediction performance for specific variables of interest or particular geographic regions. Ultimately, this has potential in creating bespoke models for weather agencies without the ability to access the large compute required for NWP. 

Weather prediction and patentability

The forecast for patenting weather prediction initially appears gloomy, but may be brighter with further consideration. 

The algorithms and mathematical methods used in weather prediction would likely fall foul of UK and European law on patentability. In T 1798/13, the EPO Board of Appeal decided that “the weather is not a technical system which can be improved, or even simulated with the purpose of trying to improve it. This kind of modelling is rather a discovery or a scientific theory” [6]. 

However, as examples, use of particular sensor data, or applications of a weather prediction system (e.g., in controlling an energy generation system, or in providing output signals or warnings), may be protectable by patent in the UK and Europe. Moreover, other jurisdictions may be more accommodating of patent applications for weather forecasting.

Summary

The use of AI for weather prediction is developing at an impressive rate, and in the process is helping to facilitate the green energy transition. Moving forward it will be interesting to see the continued development of these weather prediction models and how parties in the field look to protect these innovations.

[1]https://marks-clerk.page/ai-report-2024#ai-in-green 

[2]https://www.metoffice.gov.uk/about-us/who-we-are/innovation/supercomputer

[3]https://deepmind.google/discover/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/

[4]https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10319675 

[5]https://arxiv.org/pdf/2404.00411

[6]https://www.epo.org/en/boards-of-appeal/decisions/t131798eu1