Federated Learning and Crowdsourced Weather Data: Practice and Experience

In the era of advanced meteorological data platforms such as Copernicus and Climate Data Store, the frontier of weather forecasting has evolved. The primary challenge is no longer the acquisition of accurate and high-resolution data, but rather the effective integration and utilization of diverse observational datasets to enhance localized weather predictions. Crowd sensed weather data through a network of low-cost, widely distributed weather stations can provide the granular data needed for precise local forecasts. However, this approach introduces challenges such as data integration, consistency, and privacy concerns. Federated Learning (FL) addresses these issues by enabling decentralized data processing while maintaining data privacy.This paper introduces an innovative implementation of a federated learning framework integrated with a cluster of Automated Weather Stations (AWS). The primary objective of this study is to leverage federated learning to enhance the predictive accuracy of the Weather Research and Forecasting (WRF) model by using each weather station not only as a data acquisition point but also as a computational node. This decentralized approach maintains data privacy and security while enabling local training of models, such as Crossformer, Autoformer, and DLinear. These models' locally trained weights are periodically aggregated on the central server, which updates and redistributes the global model.Based on data collected over two years from two automated weather stations, the experimental results analyze the possibility of improving WRF model predictions for temperature and humidity. This research highlights the potential of Federated Learning in meteorological applications, offering a robust solution for enhancing weather forecast accuracy while ensuring data privacy and efficient resource utilization.