Federated Learning Framework for Privacy-Preserving Global Weather Forecasting

Countries around the world rely on heterogeneous observations and reanalysis data disseminated by national and regional meteorological agencies at the time of weather forecasting. Cross-border sharing of data may, however, be restricted by sovereignty matters, security, commercial and regulatory matters. The authors of this paper propose a federated learning framework Fed-WeatherNet to allow participating countries to predict weather forecasts on global scale while retaining full privacy control over their atmospheric data. Instead of sharing atmospheric data, the participating countries just share privacy-preserving updates. The proposed framework used an STGN or Spherical Transformer backbone combined with Climate-Cluster Weighted Aggregation (CCWA) for climate-dependent non-IID data. Information leakage can be limited through the incorporation of client-side gradient clipping, differential privacy using DP-SGD and secure aggregation. Further protection is optionally provided by homomorphic encryption which offers a privacy mechanism comparable with secure aggregation. To lower the communication cost to bandwidth constraint participants, we introduce top-k sparsification, quantization, and asynchronous participation. A standardized evaluation protocol defines centralized as well as localized-only and federated baselines, with latitude-weighted RMSE, ACC, and CRPS as main forecast metrics. The specifics of experimentation are also covered in these Non-IID robustness, privacy-utility trade-offs, communication efficiency, participation fairness, adversarial robustness, and simulated cross-country ERA5 deployment.