Evaluation of Data-Driven Medium Range Weather Prediction Using Reanalysis Data

Timely and accurate weather forecasts are essential for decision making regarding public safety, water and weather infrastructure planning, as well as energy and environmental resources management and allocation. Forecast products (e.g., wind speed/direction, temperature, and precipitation) derived from the operational numerical weather prediction (NWP) models such as the National Oceanic and Atmospheric Administration (NOAA) High Resolution Rapid Refresh (HRRR) model, Global Forecast System (GFS), and the European Center for Medium-Range Weather Forecasts (ECMWF) empower users to make informed decision and address various challenges pertaining to weather-related hazards. Despite their importance, numerical weather prediction (NWP) models face challenges in effectively representing the changing atmospheric dynamics and leveraging the increasing volume of available weather and climatological data. With vast archives of observational data at hand, traditional NWP models have struggled to fully exploit this information to enhance forecast quality. In contrast, data-driven deep learning models have great potential to improve weather predictions by addressing inherent model biases within the NWP models and facilitating the creation of extensive ensembles for probabilistic forecasting. Leveraging reanalysis data or observations during training, data-driven models circumvent limitations present in NWP models, such as biases in convection parameterization schemes that strongly affect precipitation forecasts. Moreover, data-driven models exhibit remarkable efficiency, boasting order-of-magnitude faster inference times compared to traditional NWP models. This enhanced speed allows for the generation of large ensembles, thereby providing valuable probabilistic information for decision-making processes. In this paper, we develop a transformer-based deep learning model that can provide 10-day global forecasts with a temporal resolution of 6 hours and a spatial resolution of 0.25°, providing predictions including 5 upper-air atmospheric variables at 13 pressure levels and 5 surface variables. Specifically, the 5 upper-air atmospheric variables include geopotential, temperature, wind (both u and v components), and relative humidity, and the 5 surface variables include 2-m temperature, 10-m wind (both u and v components), mean sea-level pressure, and total precipitation. Quantitative evaluation with reanalysis data and other observations have demonstrated the encouraging performance of this machine learning-based weather forecast scheme.