The part Chinese AI could play in predicting extreme weather

On 9 August, Typhoon White Dolphin arrived on China’s eastern coast, breaking rainfall records as far inland as Henan province. One weather station in the county of Lushan saw more than half its annual average rain fall in 24 hours. Flood warnings were issued and trains halted.

Sudden localised events like this are hard to predict. In August last year, a flash flood in the usually arid Yuzhong, Gansu province killed 32 people and caused CNY 2.5 billion in damage. The weather system developed rapidly over a small area, causing far more intense rain than either conventional or AI models had forecast. A government report on the disaster stated that accurate predictions will require filling in weather monitoring blind spots in remote mountainous areas.

There are high hopes for AI forecasting, which is cheap and fast – and an area in which Chinese models appear to be in the lead. Various companies have launched models while the China Meteorological Administration has an early-warning system, Mazu, being used in over two dozen countries.

However, questions remain. How accurately can these models predict extreme weather? What are the perils of relying on AI alone? Is energy use an issue? Dialogue Earth spoke to the experts.

China’s AI weather forecasting rush

Traditional weather forecasting relies on numerically solving equations of atmospheric physics, which can take several hours even on a supercomputer. AI modelling, though, reanalyses conventional weather forecasting data and uses it to make predictions. The whole process takes a matter of minutes and a thousandth of the compute power.

It seemed to take off in China in 2023. In July, the forecasting model Fengniao bettered conventional models when it predicted the path of Typhoon Doksuri, with an error of only 38.7 km, 24 hours in advance. Fengniao had been developed at the Shanghai Artificial Intelligence Laboratory by AI company Xiangfeng Technology, two universities and the Chinese Academy of Sciences.

Meanwhile, Nature published a paper on Pangu-Weather, developed by Huawei subsidiary Huawei Cloud. And Fuxi, produced by Fudan University and the Shanghai Academy of AI for Science, predicted global weather patterns 15 days in advance with relative accuracy.

It wasn’t long before Chinese government agencies got involved. In June 2024, the China Meteorological Administration published three models, one of which could predict 60 days of global weather in three minutes. In July 2025, the agency released Mazu, a global early warning system combining satellite data and AI modelling.

Visitors encounter the weather warning system Mazu at the China Meteorological Administration booth during the 2026 World AI Conference in Shanghai (Image: Sipa US / Alamy)

Some local governments were ahead of the game, said Yuan Xingyuan, founder of Colourful Clouds Tech.

His was one of the first AI companies, founded in 2014, to work on minute-by-minute precipitation forecasting. Initially it focussed on local forecasts in the coming hour, then expanded into providing flash flood early warnings to government. It has also supported a forest-fire training exercise in Liangshan, Sichuan.

“At first it was grassroots government agencies using new tech to avoid and mitigate disasters and predict flash floods,” Yuan Xingyuan said. When those agencies saw how useful the technology could be, cooperation continued.

“We’ve been providing technical support for flash flood prediction in Liangshan for years. That led to the transportation and water authorities recognising the value of our work. Over time, government agencies have come to rely on our prediction technology.”

How accurate is AI?

Extreme weather events come in many forms: heatwaves, cold waves, heavy precipitation, drought, tornadoes, tropical cyclones. All are complex systems with little modelling data to work with.

“Currently, AI global weather forecasting models do particularly well on mid-range, large-scale atmospheric circulation, and on some extreme temperature events and typhoon paths,” Zhang Wei, director of Xiangfeng Technology, told Dialogue Earth. “In certain tests, some models are already beating advanced conventional forecasts. However, the models have limitations when it comes to forecasting typhoon intensity, severe convective weather [which can cause storms] and extreme precipitation,”

“Accurate prediction of convective weather is hard anywhere. These are localised, quickly developing non-linear occurrences. The further ahead you’re looking, the harder it is to be sure about location and intensity,” he said. In the Yuzhong case, extreme rain broke multiple records and was far in excess of what had been forecast by AI or conventionally.

Rescuers clear debris from a disaster-affected house in Yuzhong, Gansu province (Image: Hou Chonghui / Xinhua / Alamy)

Current AI models have a weakness when it comes to predicting extreme weather: if it isn’t in their datasets, they can’t predict it happening. In a paper testing popular AI weather models, Zhang Zhongwei, a post-doctoral researcher at Germany’s Karlsruhe Institute of Technology, found they didn’t match up to the most advanced conventional models when predicting record-breaking events, systematically underestimating frequency and intensity.

But another paper, published in 2025, found an AI model could have successfully predicted the Dubai floods of 2024 eight days in advance. It used translocation, taking a model from a data-rich location and applying it to the Dubai region, where historical data is lacking, to predict extreme events.

Zhang Zhongwei says translocation has a lot of potential. But he warns the improvements will be limited. What is needed is more and better data.

Yuan Xingyuan also stressed the importance of data quantity. “The most accurate model will be the one with the most data.” He thinks that even low-quality data sets should be included in model training, but marked as such.

Academics think both conventional and AI-powered models need large quantities of granular, accurate, representative and diverse data on which to base predictions. That means remote places like Yuzhong, or parts of countries in the Global South, can easily become blind spots, with the lack of historical data hindering the accuracy of forecasts.

Zhang Wei told Dialogue Earth that low-cost automatic monitoring stations could be used in the Global South when the quantity or quality of data is lacking. The cost of these varies with the accuracy and quality of the data collected, but the cheapest can be somewhere in the tens of thousands of yuan (or thousands of USD). Large weather forecasting models usually rely on direct sources of primary data, such as ground-based radar, satellites and weather monitoring stations, but data from those automatic monitoring points can also be used.

Alongside the compute cost advantages, large forecasting models can also be trained on data from specific countries to meet local forecasting needs. At the World Artificial Intelligence Conference, Furrukh Bashir, head of research at Pakistan‘s Meteorological Department, gave the example of combining local satellite monitoring data with the Mazu platform to create custom storm-tracking technology. That allows for earlier warnings during Pakistan’s flood season, meaning damage can be reduced.

Credibility and energy use

Although there is general optimism about the prospects for AI forecasting, doubts remain. One point of controversy is the “black box” nature of the algorithm – the ways in which the models produce their results are not understood and cannot be tracked.

“Further discussions are still needed regarding the verification procedures and standards for AI-based weather forecasting”, said Shipra Jain, an assistant professor at University College London, at London Climate Action Week in June.

The conventional forecasting models may be more expensive, but at least their errors can be tracked and explained. But with different countries competing on AI, the issue may lose out to efficacy. At the London climate week, AI meteorology researchers expressed worries that engineers in the industry are focused on speed and better assessment results, overlooking other goals such as interpretability, reliability and justice.

In response, Zhang Zhongwei said the accuracy of these models needs to be verified with more diverse and comprehensive testing before being used in high-stakes applications like early warning systems and disaster management. A related study suggests first developing interpretability in order to boost credibility and transparency

The other elephant in the room is the energy used to train these models.

A recent study found that training a Google weather model used as much electricity as eight or nine UK households would in a year. Data storage, data centre cooling and day-to-day operations all use energy.

That said, within 17 days of use, that training energy use had been matched by what is needed to power a supercomputer for conventional forecasting. And within a year, the study found data-driven models are estimated to consume at least 21 times less energy than the physics-based model.

The future of weather forecasting

Experts consulted said we won’t be able to rely on AI weather forecasting alone.

A 2025 review found that cooperation between AI researchers, meteorologists, climate scientists and policymakers will be crucial for improving how AI is used for extreme weather events. AI can support existing infrastructure and expertise, rather than replacing them, says the World Meteorological Organisation.

Zhang Wei added that AI forecasting is already filling in gaps in conventional forecasting – but experienced meteorologists still need to check its output and make final decisions.

More accurate forecasts are only the first step to climate adaptation and disaster reduction. For example, in the August 2025 floods in Yuzhong, even after the warning was triggered, evacuations were too limited.

Technology progression such as AI weather forecasting, and physical infrastructure adaptations, such as levees or sponge cities, will never be enough. As Dialogue Earth has reported, emergency-response systems, risk management and multiagency cooperation are also vital.