Hybrid meteorological forecasting: ML-driven predictions of Lake Michigan's lake-effect precipitation for urban preparedness

The "lake-effect" is the weather trend that happens when cold air, which usually comes from Canada, moves across the open seas of the Great Lakes, warms up, evaporates as clouds, and produces precipitation. Abundant precipitation and snow have significant adverse effects on the regions surrounding the Great Lakes in North America. Traditional meteorological studies utilize fluid dynamics to model future weather and make predictions. Despite their capacity to generate precise predictions regarding numerous climate variables, fluid dynamic models encounter limitations in producing consistent and reliable forecasts over an extended temporal scope. Given the swift advancements in deep learning and time series forecasting, we have endeavored to leverage statistical modeling while increasing the typical short-term prediction range from 24 hours to a typical midrange of 3 days, with a special emphasis on severe weather events. We establish a new data pipeline that extracts visible band satellite data from the Lake Michigan region, sourced from the Geostationary Operational Environmental Satellite under the aegis of the National Aeronautics and Space Administration. Subsequently, the study introduces a framework that amalgamates 2-dimensional satellite imagery and 1-dimensional meteorological data, even when they differ in sampling frequencies and availability, for the purpose of forecasting. By integrating Convolutional Neural Networks and Long Short-Term Memory Networks, our framework not only ensures swift implementation and rapid training but also remains efficient in terms of computational expenses. Data left out of the training pipeline are used to measure the framework's accuracy in forecasting. This research contributes to the existing scientific dialogue concerning lake-effect snow and paves the way for more interdisciplinary approaches to climate change.--Author's abstract

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.