FengQing

FengQing is a data-driven model for medium-range weather forecasting jointly developed by the China Meteorological Administration and School of Software at Tsinghua University. It leverages advanced machine-learning techniques to enhance weather forecasting and analysis for over 10 days. The runnable model and inference code are fully open-sourced, allowing users to set up and perform their weather predictions efficiently.

Version

CMA-AIM-GFS-Fengqing-V1.5Beta

• 6h total precipitation added

Data and Model Download

To begin working on this project, you need to download the sample input data and pre-trained models from this zenodo link or Google Drive, which is organized according to the following directory structure:

project_root/
├── data/
│   ├── YYYY
│   	├── MMDD
│   		├── pressure_060000.npy
│   		├── pressure_120000.npy
│   		├── surface_060000.npy
│   		├── surface_120000.npy
├── mean_std/
│   ├── upper_mean.npy
│   ├── surface_mean.npy
│   ├── upper_std.npy
│   ├── surface_std.npy
│   ├── res_upper_std.npy
│   ├── res_surface_std.npy
├── utils/
│   ├── constant_masks.npy
│   ├── cfg_15days.pt
├── onnx/
│   ├── weights.pb
│   ├── fengqing.onnx

You can customize the input data by modifying get_data function in inference_fengqing.py. To ensure compatibility, please keep the exact input variable order and units as specified in the [details of data](#Details of Data).

Quick Start

Once downloading the model and sample data, follow these steps to install the necessary environment and run the inference script. When using onnxruntime-gpu, the model requires 26.5 GB of GPU memory to run effectively. Please ensure you have a GPU with sufficient memory for optimal performance.

1. Install dependencies

   pip install -r requirements.txt

2. Run Inference (start time corresponds to the last input frame)

   python inference_fengqing.py --dataset-path /project_root/data --datetime start_time --output_dir /path/to/output

Details of Data

The model takes in two main data types: upper-level and surface data. Each has its own shape and order, and any customized data should be preprocessed in the following format.

• upper-level data: [variable, level, lat, lon] rearrange to [variable * level, lat, lon]

upper variable	abbreviation	unit
Geopotential	Z	$\text{m}^2/\text{s}^2$
Specific humidity	Q	$\text{kg}\ \text{kg}^{-1}$
Temperature	T	$\text{K}$
U component of wind	U	$\text{m/s}$
V component of wind	V	$\text{m/s}$

The pressure levels are arranged in the following sequence:

1000 hPa, 925 hPa, 850 hPa, 700 hPa, 600 hPa, 500 hPa, 400 hPa, 300 hPa, 250 hPa, 200 hPa, 150 hPa, 100 hPa, 50 hPa.

• Surface data: [variable, lat, lon]

surface variable	abbreviation	unit
Mean sea level pressure	MSLP	$\text{Pa}$
10 metre U wind component	U10	$\text{m/s}$
10 metre V wind component	V10	$\text{m/s}$
2 metre temperature	T2M	$\text{K}$
Total Precipitation (optional)	TP	$\text{mm}$

Update

​ We've successfully uploaded the total precipitation model, which is now available for use on Google Drive. To fully reproduce the results, please obtain the total precipitation data for a 6-hour period (in mm units) and concatenate it with the surface input. The mean and standard deviation values required for preprocessing can be found in this Google Drive link.