Federated Learning for Global Carbon Emission Forecasting: A Hybrid Time-Series Approach with Statistical and Neural Models

The paper introduces a federated hybrid forecasting framework for global carbon emission forecasting that combines ARIMA, GARCH, LSTM-Attention, and XGBoost within a privacy-preserving federated learning setup. Experimental results from 14 clients show R2 values ranging from 0.50 to 0.97, RMSE values between 0.06 and 2.35, and MAPE values from 1.5% to 11.3%, demonstrating the framework's accuracy and scalability. This approach is significant for practitioners as it allows for effective carbon emission predictions while maintaining data privacy and compliance with regulations.