We’ve come a long way in predicting the weather. Once, forecasts were limited to just a few days ahead, but now five- to seven-day predictions are often accurate. AI weather forecasting has become highly reliable for many standard predictions, often matching or surpassing traditional physics-based models in both speed and accuracy.
Systems like ECMWF’s AIFS and NOAA’s AI-driven models can deliver 10-day global forecasts in minutes, thousands of times faster and more energy-efficient than conventional supercomputers, while performing better on metrics like temperature and wind in roughly 90% of cases. New Zealand’s Cascade AI supercomputer has even boosted five-day forecast accuracy to match two-day levels, cutting flood model times from days to minutes. This technology is crucial as extreme weather events become more frequent.
However, AI forecasting tools still face limitations, especially with rare or unprecedented events. Extreme floods, record storms, and unusual climate conditions often fall outside AI training datasets, where physics-based models (numerical weather prediction, NWP) can still outperform AI in extrapolating beyond known patterns. Their black-box nature also makes decisions hard to interpret, and coarse resolution limits small-scale detail, such as localized storms.
AI is now being applied beyond traditional weather to areas like wildfire prediction. Researchers at the University of Canterbury have developed an AI-based wildfire forecasting system that updates every 30 minutes, compared with standard fire danger indices that refresh only once per day. Using over 60 years of historical weather and fire data, the system improved prediction accuracy by 10–30%, detecting dangerous fire conditions earlier and reducing false alarms. The model works by analysing weather patterns from existing station networks, including temperature, wind, humidity, and dryness indicators, to identify conditions that often precede fire ignition.
Using a cost–benefit approach, the researchers found that the AI system could potentially double the economic benefits of current forecasting methods by minimizing both overlooked fires and false alarms. Moreover, because the model uses data from existing weather stations, it can be implemented broadly without the need for additional infrastructure. Dr Ardid notes that this method could also be applied in New Zealand, where comparable meteorological monitoring networks are already in place, making it suitable for forestry management, regional fire agencies, and other sectors.
“The models rely on weather station data, which already exists through monitoring networks in New Zealand. That means the system could potentially be implemented without new infrastructure, either at a regional level for fire management agencies, or also at a more local or sector level, for example within the forestry industry.”
By leveraging existing weather station networks, the system could be widely deployed without new infrastructure, making it suitable for New Zealand’s forestry and fire management agencies. Economic modelling also suggests the AI approach could double the value of fire forecasts, reducing costs from missed fires and unnecessary alarms.
The combined progress in AI for weather and fire forecasting demonstrates both promise and caution. While AI excels in speed, medium-range predictions, and pattern recognition, it remains vulnerable to rare, extreme events and unexpected climate anomalies. These tools are designed to complement, not replace, traditional forecasting and human decision-making. Still, the rapid advances show that, with careful integration, AI can become a life-saving tool in an era of increasingly unpredictable weather and wildfire risk.
SPONSORED
