Climate and weather forecasts: how AI is helping with predictions

The quality of weather forecasts has already improved greatly in recent decades. One key factor in this is satellite data, which has been included since the 1980s. At the same time, more and more weather stations are being set up worldwide to provide important data. New sensor networks, including crowdsourced data from weather stations operated by private individuals, are supplementing professional measuring stations. Innovative technologies such as weather drones can collect targeted data from different layers of the atmosphere.

Forecasts for just seven days – or longer?

Although the data situation has improved significantly, today's calculation models can only predict about seven days. Anything beyond that cannot be reliably predicted with the models currently available.

However, the seven-day limit could well change in the next few years. Forecasting models that use AI technologies can process significantly more data in significantly less time. AI-based weather models could not only include all currently available data, but also take into account data from the past. This will lead to new forecasting models that recognise patterns and calculate probabilities for the resulting weather conditions.

A glimpse into the future of AI weather forecasting is provided by a development from Google subsidiary Deepmind, whose AI model is said to be able to deliver ‘faster and more accurate forecasts for up to 15 days’. The AI ‘GenCast’ was trained with weather data from the last four decades and generates a probabilistic weather model that is more accurate than previous models.

In initial tests with real weather data from 2019, GenCast provided better forecasts than the current top model of the European Centre for Medium-Range Weather Forecasts (ECMWF) in 97.2 per cent of cases. Experts see this as an important first step, but there are still problems to be solved.

AI forecasts require real-time weather data, preferably worldwide. With several thousand weather stations around the globe, this is a logistical challenge. It is questionable whether the computing power of today's data centres is sufficient for this. It is possible that AI forecasts will initially be limited to a maximum of 15 days until computing power increases. However, with the first quantum computers, this could soon be the case.

Climate calculations are even more complex

How will the climate change if the earth continues to warm? What effects will people in different regions have to prepare for in the coming years? To answer these questions reliably, climate researchers need weather data that is as complete as possible over very long periods of time.

While sufficient data is available for weather forecasting, this is not the case for climate calculations using machine learning and AI. A complete data set has only existed since around the middle of the 20th century. The first task of AI models is therefore to reconstruct weather data in order to obtain complete data sets for a longer period of time.

For example, a study by an international research group led by Étienne Plésiat and Christopher Kadow from the German Climate Computing Centre (DKRZ) used artificial intelligence to reconstruct past extreme events and reveal spatial trends over a long period (1901-2018) not covered by most inventory data.

Only when the historical context has been sufficiently determined can AI models contribute to our understanding of current climate extremes and future climate risks.

Outlook on tomorrow's forecasts

The outlook for future developments is promising. Quantum computing could enable even more complex calculations in the future. New generations of satellites will deliver even more precise data. The further development of AI will lead to even more intelligent forecasting systems. At the same time, international cooperation in weather research is becoming increasingly important to better understand and predict global phenomena. These advances will not only improve the accuracy of forecasts, but also extend their time horizon.

Text: Falk Hedeman