With the increasing urgency for the UK to strengthen its resilience to extreme weather events, Dr Simon Driscoll and Dr Natalie Harvey from the University of Reading looks at the AI forecasting revolution and the mixed ability of this approach to capture a record-breaking windstorm
In recent years, we have been witnessing a revolution: the rise of artificial intelligence (AI) and machine learning (ML).
AI and ML methods have now gone vastly beyond the early ‘perceptron’ model (developed to mimic neurons in brains) to finding applications in economics, science, technology, industry and more.
It is now common to ‘talk’ to AI chatbots such ChatGPT and Gemini to help with our daily lives, and the terms machine learning and AI are now ubiquitous.
ML and AI refer to similar things: a set of algorithms that can learn from and make predictions on data, with AI also more broadly referring to technologies using these algorithms for robotics, problem-solving and so on. These statistical algorithms thrive off big data.
ML and AI forecasting
Forecasting the weather is a computational problem described as being akin to simulating the human brain and the evolution of early universe. So it is natural that ML and AI techniques have found themselves to being applied in this area.
The current state of the atmosphere is found by combining information from a vast network of observations that include data from weather stations at the surface of the Earth right through to data from satellites in space (and everything in between!).
This data is then fed in to as a starting point for the numerical weather forecasting models. These forecast models use the fundamental laws of physics to predict the state of the atmosphere days and weeks in ahead.
The atmosphere is a chaotic system, meaning that two atmospheres almost identical in their state, could rapidly diverge to have different weather conditions for the coming days. Therefore, forecasting the weather reliably requires vast computations on supercomputers that require regular updates from this observation network.
Recall that ML and AI algorithms learn patterns from data and make predictions from that data. Thus, if ML and AI algorithms could learn from the vast amount of ‘reanalysis’ data (data that represents the best estimate of the weather at that time) maybe they could learn to forecast the weather.
While ML and AI have been applied in climate contexts before, the first attempt to use these techniques to forecast the weather came in 2018 by scientists at the European Centre for Medium Range Weather Forecasts (ECMWF).
This work was seminal in itself but it did not rival at the time traditional physics-based numerical weather prediction.
This has all changed in the last few years. Advancements in ML techniques, such as using graphics processing units (called GPUs) to perform calculations, alongside the publication of a WeatherBench dataset, a 10-year roadmap for ML by the ECMWF, and other developments, have changed the landscape dramatically.
Enter the tech giants: in 2022 and 2023 four ML/AI models for short-to-medium range weather forecasts were developed by Huawei, NVIDIA and Google.
By combining the world-leading resources from ECMWF for the training data with substantial resources and skills available to the tech giants, the world has witnessed a rapid increase in the skill of ML and AI models.
In many domains the AI models can now produce more skilful forecasts than numerical weather prediction (NWP) models, and also run substantially faster than traditional models.
Storm Ciarán
In late 2023, Storm Ciaran evolved from a weak disturbance in the North Atlantic to a storm with a record-breaking low pressure centre.
It landed on French and British shores in early November bring winds with gusts of over 115 miles per hour. The strong winds killed at least 16 people, shutdown schools and airports and over 1m homes were without electricity.
A group at the University of Reading, led by Professor Andrew Charlton-Perez, sought to understand how well these AI forecasting models could model this storm, and how they compared with existing NWP models.
Storm Ciarán is an out-of-sample test for the AI models at the time of the study – meaning that the models had not been trained on its data. If they could predict the features accurately of this storm, it would lend confidence to AI models being used and able to predict other storms in the future.
Comparing the AI models with NWP
The study analysed the performance of the AI models of Huawei, Google and NVIDIA and the NWP models of the ECMWF, UK Met Office, Japan Meteorological Agency and National Centers for Environmental Prediction (USA).
All models tracked the evolution of the storm across the Atlantic and captured well the mean sea level pressure low corresponding to the centre of the storm.
However, perhaps surprisingly given the incredible march of AI in the recent years in the weather forecasting domain, the AI models failed to capture the peak wind speeds seen in Storm Ciarán, while the NWP models fared substantially better.
This is significant for two reasons. Firstly, weather warnings are based on forecast wind speeds being over a certain threshold. Based on the wind speeds predicted by the NWP models, a red risk to life warning was issued in areas of the UK and France.
Potentially, if the AI forecasted winds were used, a lower category of warning could have been issued, possibly leading to a higher death toll.
Secondly, small differences in wind speed can lead to large differences in the estimates of economic loss associated with a windstorm as they are generally modelled as the cube of the normalised wind gust speed.
The team also examined the dynamical structure of Storm Ciarán by comparing the forecasts to data from ECMWF, giving the best estimate of the evolution of the storm (known as reanalysis data).
While many important dynamical features of Storm Ciaran were captured well by the dynamical models, they failed on others – such as the strength of the temperature changes across the weather fronts, and again failing to produce the narrow band of wind that caused the most severe impacts.
Future of AI and UK resilience to extreme weather events
Despite the tremendous success of AI models so far, the study highlighted potential flaws for these models – in particular in terms of capturing features important for damaging extremes.
In a report on extreme weather risks, the House of Commons Public Accounts Committee raised concerns about UK government’s approach in being able to strengthen UK’s resilience to society-wide risks (including extreme weather risks), saying it “lacks the required robust leadership, oversight and urgency”.
With an increase in the likelihood of extreme weather events as climate change continues, the need to cope better with these risks will be of greater importance.
Of particular relevance, Storm Ciarán was not an unusual storm, and as part of UK preparedness, further work must also go into understanding how much the AI models’ predictions can be trusted for more dynamically unusual storms.
The use of AI models in forecasting the weather and such extremes will also only increase – the ECMWF now have their own AI model rivalling the tech giants (called AIFS).
This highlights the imperative to understand what these models are physically doing, explaining how and why in some respects they are doing so well and why they are missing certain features.
Careful communication must occur as to their strengths and weaknesses to the public, decision- and policymakers and the insurance industry – especially in the case of extreme events, to avoid an unnecessary loss of lives and weather-related damage.
With a correct understanding and improvement of these models, building on studies like this through further research, the already witnessed revolutionary potential of these AI models means that AI weather forecasts of many thousands of realisations with great accuracy, speed and physical consistency might not be too far away.
Dr Simon Driscoll
University of Reading
Dr Natalie Harvey
University of Reading
