Subseasonal Predictability of Diverse Persistent Wet-Cold Compound Extreme Events over China Based on Calibrating Dynamical and AI models
This study evaluates and calibrates dynamical and AI-based subseasonal models to demonstrate that persistent wet–cold compound extreme events over China are predictable up to 15–20 days, with the Central China type showing the highest skill driven by specific large-scale circulation patterns, while highlighting distinct performance strengths between the ECMWF and CMA-FENGSHUN systems across different lead times.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Every winter, parts of China face a particularly punishing kind of weather: a persistent combination of heavy rain or snow and biting cold. Scientists call these persistent wet–cold compound extreme events. They are more than just a cold snap or a rainy day; they are the dangerous overlap of both, creating conditions like freezing rain and heavy snowfall that can paralyze cities, damage power grids, and threaten lives. While we know these events happen, predicting exactly when and where they will strike more than a week or two in advance has been a stubborn challenge. This is the "subseasonal" gap in forecasting—the awkward time between a standard weather report, which is reliable for a few days, and a seasonal outlook, which looks at months ahead. In this window, the atmosphere is too chaotic for short-term precision, yet too fast-moving for the slow changes of the ocean to provide a clear guide.
A team of researchers from the China Academy of Meteorological Sciences and the National Climate Center has taken a fresh look at this problem. They focused on three distinct types of these wet–cold events that occur in different regions of China: one affecting the south, another the west, and a third the central plains. To tackle the difficulty of prediction, they did not rely on just one method. Instead, they compared traditional computer models, which simulate the physics of the atmosphere, against a newer generation of artificial intelligence systems that learn patterns from vast amounts of historical data. Crucially, they first applied a statistical correction to the raw data from these models. Think of this step like tuning a radio to remove static; the models often have built-in biases, such as predicting rain that is too light or temperatures that are too warm. By adjusting the forecasts to match the real-world distribution of weather, the researchers could see the true signal of the models without the noise of their errors.
The results of this comparison reveal a promising path forward. After the statistical tuning, all the models—both the physics-based ones and the AI system—became significantly better at identifying where these dangerous events would occur. They could reliably map out the general shape and location of the wet–cold zones up to two weeks in advance. However, the study found that not all events are created equal, and not all models are equally good at every stage. The events in western China, which often form over the high Tibetan Plateau, were the most consistent to predict in terms of their continuous weather patterns. In contrast, the events in central China were the most predictable overall, largely because they are driven by very organized, large-scale waves of air moving across the globe that are easier for models to track.
When it comes to the models themselves, a clear pattern of strength emerged based on time. The European Centre for Medium-Range Weather Forecasts model, a traditional powerhouse, performed best in the first few days, offering sharp and accurate short-term warnings. But as the forecast extended further out, into the second week, the AI-based model, known as FENGSHUN, began to pull ahead. It held onto its predictive skill longer than the others, maintaining a clear picture of the wet–cold anomalies even at the 15-to-20-day mark. This suggests that while traditional models are excellent at capturing the immediate physics of the atmosphere, the AI system has learned to recognize the deeper, longer-term patterns that drive these events, even if it does so without explicitly calculating the physical forces at play.
The researchers also dug into the "why" behind these predictions. They found that the most predictable events were always preceded by specific, robust signs in the atmosphere. For the central China events, a clear wave train of high and low pressure stretched from Europe across Asia, accompanied by a distinct suppression of cloud formation over the tropical oceans. When these large-scale signals were strong and organized, the models could forecast the event with confidence. When the background atmosphere was messy or the signals were weak, the predictions fell apart. This distinction is vital: it means that the ability to predict these disasters depends heavily on the state of the atmosphere before the event begins. The study concludes that while we cannot yet predict every single wet–cold event with perfect certainty, we now have a much clearer understanding of which ones are predictable and how different tools can be used together to extend our warning time, potentially saving lives and reducing economic losses in the weeks before disaster strikes.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.