Prompt Federated Learning for Weather Forecasting: Toward Foundation Models on Meteorological Data

This is the official repository for the paper Prompt Federated Learning for Weather Forecasting: Toward Foundation Models on Meteorological Data Accepted by IJCAI'23 .

Abstract

To tackle the global climate challenge, it urgently needs to develop a collaborative platform for comprehensive weather forecasting on large-scale meteorological data. Despite urgency, heterogeneous meteorological sensors across countries and regions, inevitably causing multivariate heterogeneity and data exposure, become the main barrier. This paper develops a foundation model across regions capable of understanding complex meteorological data and providing weather forecasting. To relieve the data exposure concern across regions, a novel federated learning approach has been proposed to collaboratively learn a brand-new spatio-temporal Transformer-based foundation model across participants with heterogeneous meteorological data. Moreover, a novel prompt learning mechanism has been adopted to satisfy low-resourced sensors' communication and computational constraints. The effectiveness of the proposed method has been demonstrated on classical weather forecasting tasks using three meteorological datasets with multivariate time series.

Code guideline

Path of the proposed SPL: base_module/pretrain_trans.py/[Novel_Prompting]

Path of the framework optimization: GraphGenerator.py, aggregator.py

Datasets and experiments

All of the dataset utilized in this paper can be found in National Aeronautics and Space Administration (NASA)

Simple Instruction for Dataset Processing: (Dataset) To use the datasets in our paper, first select a subset from the NASA-provided dataset based on the client number and latitude/longitude ranges detailed in the Appendix of this paper. Then, process it in array format. Self-processing ensures that the data is appropriate for your particular research scenario and complies with policies regarding data use and sharing.

Please cite our publication if you found our research to be helpful.

@article{chen2023prompt,
  title={Prompt Federated Learning for Weather Forecasting: Toward Foundation Models on Meteorological Data},
  author={Chen, Shengchao and Long, Guodong and Shen, Tao and Jiang, Jing},
  journal={arXiv preprint arXiv:2301.09152},
  year={2023}
}

Contact

If you have any questions, please do not hesitate to contact me (Email: shengchao.chen.uts@gmail.com).