AI Enhances Snowfall and Wind Forecasting Accuracy

The University of Connecticut researchers are applying artificial intelligence (AI) and machine learning (ML) to enhance the predictability of snowstorms, particularly extreme weather in the Northeast United States. Associate Professor Marina Astitha and the UConn Atmospheric and Air Quality Modeling Group are applying AI to conventional and physics-based models to modify the forecasting of snowfall accumulation and wind gusts. The outcome of their latest research was made public in the Journal of Hydrology and Artificial Intelligence for the Earth Systems.

The Northeast presents especially challenging forecasting conditions, especially during winter with messy weather patterns such as Nor'easters. They are wildly varying storms in size and motion patterns, and so they are hard to predict. Traditional numerical weather prediction (NWP) models predict by applying physical rules and modeling atmospheric behavior with math equations. Yet, because of the size of weather systems and the turbulent nature of the atmosphere, such models tend to be constrained by uncertainties, particularly when modeling microphysical processes such as precipitation.

To meet all these needs, the UConn research team started working on machine learning-based techniques around five years ago. They are working to complement existing weather prediction models and give improved short-term forecasts for snowfall and gusty winds. These forecasts are not just useful for public information in general but also for infrastructure design, power loss management, and emergency response.

In their latest study of snowfall forecasting, postdoc Ummul Khaira targeted the ensemble convergence in order to make it more reliable. She applied an ML platform together with Weather Research and Forecasting (WRF) model inputs to increase temporal and spatial resolution. Using multiple ensemble inputs and ML corrections, the model provides better snowfall total forecasts, especially for rapid-streets storm events that otherwise proved challenging to forecast accurately.

Another research priority is anticipating wind gusts. Israt Jahan, a graduate student, has been constructing models to enhance short-term prediction of destructive gusts. Gusts of wind are particularly challenging to predict as they differ from steady winds and can change drastically within seconds. Intense flashes of high-speed wind have the potential to produce tree damage and property loss, resulting in power outages and roadblockage. The researchers' work focuses on determining trends that are a more accurate predictor of gust intensity and timing through the use of AI-boosted models.

These innovations are being applied in partnership with the Eversource Energy Center, enabling forecasts to flow directly into operational planning and emergency response measures. Reliably predicting weather allows energy companies and emergency responders to pre-position resources before storms strike, saving on recovery time and enhancing safety.

Forecasting models such as the WRF model the atmosphere on a three-dimensional grid, numerically solving motion and thermodynamic quantities in each grid cell. Model resolution has improved, but cells remain large—usually one to four square kilometres—so they can't handle details such as micro-scale cloud or wind patterns. AI helps by capturing patterns in large sets of data that could be missed or generalized by traditional models. This enhances the precision of predictions without the requirement of more elaborate physical simulations.

The study also tests the limitations of single deterministic model output usage. One forecast only addresses one potential future state. But with the utilization of an ensemble of models and artificial intelligence-based uncertainty analysis, the team is capable of providing a range of potential outcomes that enable more empowered decision-making. This is especially useful for scenarios such as bomb cyclones, which form quickly and potentially bring sudden and dramatic changes in weather.

The research also points to increasing applications of machine learning for environmental modelling, where AI systems analyze past weather patterns and learn from forecasting errors to enhance future predictions. Far from perfection, the tools facilitate quicker processing and more precise correction of model biases that have been identified. By combining physics-based and data-driven approaches, researchers aim to develop hybrid systems that leverage strengths from each methodology.

The ultimate long-term goal for the team is to develop higher-resolution regional forecasting systems that can ultimately be deployed in several regions and under various weather conditions. Much of past AI work toward weather forecasting has been geared toward global or high-scale uses. The research being done at UConn aims to put an area of much-needed emphasis on high-resolution regional forecasting, particularly valuable for high-density or infrastructure-impact-critical regions such as the Northeast United States.

This study is nearer to filling the gap between theoretical atmospheric science and actuality. The union of large computing power with meteorological models is not only enhancing the quality of forecasts but assisting populations in reacting more effectively to dangerous weather occurrences.

Source:

Elaina Hancock, University of Connecticut. Based on research published in Journal of Hydrology (2024, 2025) and Artificial Intelligence for the Earth Systems (2024).

2025 University of Connecticut / Phys.org