This is the first time Taiping Reinsurance is bringing high-resolution flood modeling in-house, instead of relying on coarser models developed by US risk analytics companies, said Sheldon Yu, chief executive officer at the Hong Kong-based reinsurer
The weather forecasting industry has made big leaps in accuracy but has struggled with hyper-local predictions. But the proliferation of AI weather models in recent years means small, commercial firms are now developing the ability to rapidly make specialised predictions, like when and how much it will rain in your neighbourhood or how much wind will blow to spin a turbine.
For decades, public agencies have run global weather models that require supercomputers to crunch complex physics equations to spit out predictions. The need for more granular forecasts is becoming more pressing as climate change increases the odds of extreme weather, and AI is poised to offer a cost-effective way to provide them.
“The application of a previously trained machine-learning weather forecasting model, in terms of computing, costs nothing,” said Peter Bauer, a scientist at the Max Planck Institute for Meteorology and formerly at the European Centre for Medium-Range Weather Forecasts.
Among the new players seeking to capitalise on commercial opportunities in forecasting is weather tech startup Stellerus, spun out of research conducted by scientists at the Hong Kong University of Science and Technology (HKUST).
Its novel method takes rainfall forecasts, then applies a machine learning algorithm to rapidly simulate and predict potential flooding for every Hong Kong street in under three minutes. Because the forecasts can be run so quickly, flood predictions can be produced hours ahead and then updated for near-real-time warnings.
Public weather agencies “cannot do very customised predictions for a particular industry or business” despite high demand for such services, said Hui Su, chair professor at HKUST and co-founder of Stellerus. The startup also applies AI techniques to satellite data to monitor greenhouse gas emissions from specific factories and vessels.
Su’s team is working with Taiping Reinsurance, part of state-owned insurer China Taiping Insurance Holdings Co., to develop forecasts that can then be used to ping clients and policyholders with early and precise flood alerts. For example, a car owner might receive a text message to move their vehicle out of a specific garage set to be flooded in a rainstorm.
This is the first time Taiping Reinsurance is bringing high-resolution flood modeling in-house, instead of relying on coarser models developed by US risk analytics companies, said Sheldon Yu, chief executive officer at the Hong Kong-based reinsurer.
The effort comes after a series of extreme rainfall events in the past two years unleashed damaging floods citywide.
Floods are particularly hard to model because they are caused by localised weather systems and their full impact is dependent on many human-made factors, like drainage systems. But as extreme rains become more frequent, floods account for a rising share of insured losses.
Globally, more than one in five people are estimated to face significant flood risk, with South and East Asia being the most exposed. Nearly one-third of coastal residential regions of southern China’s Greater Bay Area, a cluster of nine cities that Taiping Reinsurance is focusing its flood modeling work on, is forecast to see heightened inundation threats in coming decades.
“We are pushing the industry to think outside the box of traditional ideas, and take a more proactive way to deal with catastrophe events and the losses caused to the community,” Yu said, adding that the company plans to extend its proprietary flood modeling later to Macau and Guangdong province, where the additional high-resolution flood data could shape policy discussions about establishing a province-wide flood insurance program.
Instead of relying on one global model to predict everything — like temperatures, wind speeds, rainfall, soil moisture and ozone levels — AI enables a whole new environment made up of “specialist systems that do something specific very, very well,” said Bauer.
“That’s a big, big step, in terms of what you could potentially obtain in terms of quality,” he said, adding that this creates “endless opportunities for businesses” that can produce reliable forecasts tailored to a specific customer’s needs.
Still, the growing ecosystem of commercial weather forecasters and risk modelers could mean greater dependence on proprietary models that can’t be double-checked. In an era of AI forecasting, it’s hard to trace the connection between observations, like satellite images and radar readings, and forecast accuracy, said Sarah Dance, professor of data assimilation at the University of Reading.
“You don’t really understand how changing one weight in the neural network affects the forecast very well.”
Other companies using AI to make more granular weather forecasts include Nvidia Corp. The company collaborated with the Taiwan government to develop an AI weather model trained on coarse global meteorological data that uses a machine learning technique to sharpen the resolution. The team chronicled their efforts in a Nature paper published in February.
With severe weather, “the more granular you can understand when and where events are going to occur, we believe the better reaction and planning you can make in terms of helping getting prepared for that,” said Dion Harris, senior director of high-performance computing and AI factory solutions at Nvidia.
Taiwan’s Central Weather Administration is currently evaluating the model for operational use, according to Cheng-Chin Liu, a researcher at the agency who worked with Nvidia on the project.
Meanwhile, the European forecasting body in February became the first major prediction center to launch its own AI model. The Hong Kong Observatory also references several AI models in its forecasts, including Huawei Technologies Co.’s Pangu-Weather, and is evaluating a few others, said acting senior scientific officer He Yuheng.
Even as AI weather models have gathered momentum, some meteorologists are sticking with traditional physics-based ones. In January, Swiss-based commercial forecaster Meteomatics launched a high-resolution, 1-kilometer-scale weather model for the US, following the European version’s rollout in 2022. Both are based on classical models, enhanced by complex post-processing and additional atmospheric data gathered via a fleet of in-house drones.
Running the two high-resolution Europe and US models requires over 100,000 computer processing cores, and going global “might require something in the order of 1 million,” said Meteomatics Chief Executive Officer Martin Fengler. While AI weather models may be more energy efficient, “they tend to take certain shortcuts” at the expense of accuracy and consistency, and there’s no evidence they are outperforming traditional methods, he said.
“AI is a great tool to enhance the classical way of modeling, but not fully replace it.”
Ultimately, AI weather models still depend on enormous public archives of climate and weather data compiled using physics-based methods. Most also rely on traditional prediction systems to kickstart the forecasting process. Governments should play to their strengths and continue doing the expensive work of computing and safeguarding those datasets while private companies leverage AI to make more specialized weather forecasts, said Bauer.
“The distribution of duties and responsibilities will slightly shift,” he said of the forecasting landscape. “But I don’t think it’s a problem. I think it creates more opportunities than it creates risks.”
Bloomberg