Unofficial demonstration of Huawei's Pangu Weather Model. Implementing the entire process of data preparation for input, forecasting conversion of forecasted results, and visualization.

【非官方】华为盘古天气模型演示，含输入数据准备、预测结果转换及结果可视化全流程。中文指南

1. Register for an account at Climate Data Store
2. Copy the url and key displayed on CDS API key and add them to the ~/.cdsapirc file.
3. Clone this repo and install dependencies accordingly, depending on GPU availability.

git clone https://github.com/HaxyMoly/Pangu-Weather-ReadyToGo.git
cd Pangu-Weather-ReadyToGo

# GPU
pip install -r requirements_gpu.txt
# CPU
pip install -r requirements_cpu.txt

conda install -c conda-forge cartopy

4. Download four pre-trained weights from Pangu-Weather and create a folder named models to put them in. Feel free to download only one of them for testing purposes.

mkdir models

1. Modify the date_time of the initial field in data_prepare.py.

date_time = datetime(
    year=2023, 
    month=7, 
    day=9,
    hour=23,
    minute=0)

You may check the data availability at a specific moment by using test_avaliable_time.py.You would get something like this: The latest date available for this dataset is: 2023-07-13 13:00.

2. Run data_prepare.py to download the initial field data and convert them to numpy array.

python data_prepare.py

3. Modify the following variables in inference.py according to your needs:

# Enable GPU acceleration
use_GPU = True

# The date and time of the initial field
date_time = datetime(
    year=2023, 
    month=7, 
    day=9,
    hour=23,
    minute=0)

# The date and time of the final approaches
date_time_final = datetime(
    year=2023, 
    month=7, 
    day=17,
    hour=23,
    minute=0)
# Program auto choose model to use least interation to reach final time

4. Execute inference.py to make forecast

python inference.py

5. Modify the date_time and final_date_time of the initial field in forecast_decode.py

# The date and time of the initial field
date_time = datetime(
    year=2023, 
    month=7, 
    day=9,
    hour=23,
    minute=0)

# The date and time of the final approaches
date_time_final = datetime(
    year=2023, 
    month=7, 
    day=17,
    hour=23,
    minute=0)

6. After making the forecast, run forecast_decode.py to convert the numpy array back to NetCDF format

python forecast_decode.py

7. Navigate to the forecasting directory to visualize the results

cd outputs/2023-07-09-23-00to2023-07-17-23-00
# Visualize the land surface forecast
ncvue output_surface.nc
# Visualize the upper air forecast
ncvue output_upper.nc

Don't forget to select the variable to be visualized.

Thanks Huawei team for their amazing meteorological forecasting model Pangu-Weather.
Thanks mcuntz for his/her wonderful open-source NetCDF visualization project ncvue.

I am a Bioinformatics student, not a meteorologist, so I cannot guarantee the accuracy of the code. Therefore, this project is only intended for reference and learning purposes. Additionally, this project is based on Pangu-Weather and follows its BY-NC-SA 4.0 license, and should not be used for commercial purposes. Please cite the publication of Pangu-Weather.

@Article{Bi2023,
author={Bi, Kaifeng and Xie, Lingxi and Zhang, Hengheng and Chen, Xin and Gu, Xiaotao and Tian, Qi},
title={Accurate medium-range global weather forecasting with 3D neural networks},
journal={Nature},
doi={10.1038/s41586-023-06185-3},
}

1. 前往 Climate Data Store 注册一个账号
2. 前往 CDS API key，复制url和key，写入 ~/.cdsapirc 文件
3. 克隆本仓库，根据是否有独显选择安装依赖

git clone https://github.com/HaxyMoly/Pangu-Weather-ReadyToGo.git
cd Pangu-Weather-ReadyToGo

# GPU
pip install -r requirements_gpu.txt
# CPU
pip install -r requirements_cpu.txt

conda install -c conda-forge cartopy

4. 在 Pangu-Weather 下载4个预训练模型，创建一个名为 models 的文件夹，把它们放进去（也可以根据需要任意下载一个测试）

mkdir models

1. 修改 data_prepare.py 中初始场的 date_time，

date_time = datetime(
    year=2023, 
    month=7, 
    day=9,
    hour=23,
    minute=0)

某时刻数据可用性可运行 test_avaliable_time.py 来查询.你的输出会是这样的: The latest date available for this dataset is: 2023-07-13 13:00.

2. 执行 data_prepare.py 下载初始场数据并转换为npy格式

python data_prepare.py

3. 根据需要修改 inference.py 中以下变量

# 是否启用GPU加速
use_GPU = True

# 初始场时刻
date_time = datetime(
    year=2023, 
    month=7, 
    day=9,
    hour=23,
    minute=0)
# 目标到达时刻
date_time_final = datetime(
    year=2023, 
    month=7, 
    day=17,
    hour=23,
    minute=0)
# 程式会自动选择最少到达目标时间的模型组合

4. 执行 inference.py 进行预测

python inference.py

5. 修改 forecast_decode.py 中初始场时刻和目标到达时刻

# 初始场时刻
date_time = datetime(
    year=2023, 
    month=7, 
    day=9,
    hour=23,
    minute=0)
# 目标到达时刻
date_time_final = datetime(
    year=2023, 
    month=7, 
    day=17,
    hour=23,
    minute=0)

6. 预测完成后，执行 forecast_decode.py 将npy转换回NetCDF格式

python forecast_decode.py

7. 进入预测文件路径可视化结果

cd outputs/2023-07-09-23-00to2023-07-17-23-00
# 可视化预测地表数据
ncvue output_surface.nc
# 或可视化预测大气数据
ncvue output_upper.nc

记得选择要可视化的变量

华为团队开源的气象预测大模型 Pangu-Weather
mcuntz开源的优秀NetCDF可视化项目 ncvue

本人专业为生物信息学，并非气象专业人士，无法保证代码完全准确，因此该项目仅供参考交流学习。另该项目系基于 Pangu-Weather，因此亦遵循原项目的BY-NC-SA 4.0开源许可证，切勿用于商业目的。使用本项目请引用原项目

@Article{Bi2023,
author={Bi, Kaifeng and Xie, Lingxi and Zhang, Hengheng and Chen, Xin and Gu, Xiaotao and Tian, Qi},
title={Accurate medium-range global weather forecasting with 3D neural networks},
journal={Nature},
doi={10.1038/s41586-023-06185-3},
}