Introduction
While AI Global Weather Forecasting Models (AIGWFMs) such as Pangu-weather and GraphCast have achieved performance parity with traditional numerical weather prediction (NWP) systems, they often struggle with "last-mile" operational requirements. Most state-of-the-art models are trained on the ERA5 reanalysis dataset, which restricts them to a 0.25° spatial resolution and 1-hour time intervals. This coarseness, combined with a lack of real-time local observation integration and missing outputs for critical products like precipitation or cloud cover, limits their utility for local meteorological services or aviation departments.
"YanTian" is an adaptable, plug-and-play application platform designed to bridge this gap. It wraps existing AIGWFMs in specialized modules to enable high-resolution, locally optimized forecasts and diverse product generation on consumer-grade hardware.
System Architecture
The YanTian platform employs a "loosely coupled" architecture that enhances existing models without requiring expensive re-training. It consists of four primary capability-enhancing modules:
- Local Forecast Optimization Module: This functions as a local assimilation tool. It utilizes a deep neural network to "densify" sparse, real-time station observations into grid-based corrections, providing the global model with a more accurate initial state.
- High-Resolution Forecast Module: Using a "mean-variance" approach inspired by pixel-shifting technology, this module generates 9km resolution forecasts. A conditional diffusion model is used to generate fine-grained local details (variance) grounded in the lower-resolution global trend (mean).
- Forecast Time Interval Density Module: To increase temporal resolution from 1-hour to 15-minute intervals, the platform uses bi-directional optical flow interpolation combined with local data fusion to maintain physical consistency.
- Diversified Product Generation Module: This module uses Swin-Transformers to derive 3-hour accumulated precipitation and Generative Adversarial Networks (GANs) to create synthetic satellite cloud images, features often missing from native AI models.
Data and Implementation
YanTian leverages a variety of datasets, including ERA5 (0.25° baseline), ERA5-Land (0.1° for precipitation training), real-time station data from the China Meteorological Administration, and FY-4B satellite data (4km resolution for GAN training).
A key innovation is the democratization of high-end forecasting; these modules are designed to run on a single PC equipped with a consumer GPU, such as an NVIDIA RTX 3080 or 4090. This significantly lowers the barrier to entry for small-to-medium meteorological service providers.
Key Results
The platform demonstrates significant improvements across several metrics:
- Accuracy: The local optimization module markedly reduces forecast error by integrating regional station data.
- Resolution: Spatial resolution is successfully upscaled from 25km to 9km, while temporal resolution is increased to 15-minute updates.
- Product Generation: The system generates synthetic satellite images and precipitation maps that closely align with real-world observations.
Conclusion and Future Work
YanTian provides a powerful framework for operationalizing AI weather models. While currently dependent on the quality of underlying open-source AIGWFMs, future work will focus on standardizing interfaces to facilitate the integration of new models. The ultimate goal is a collaborative, multi-party ecosystem for the continuous refinement of AI-driven meteorological tools.
Summary updated