A privacy preserving multi-center federated learning framework for district heating forecast

This paper presents a privacy-preserving Multi-Center Federated Learning (MCFL) framework for district heating demand forecasting with a 24-hour prediction horizon. To evaluate the effectiveness of this framework, we conducted a comparative analysis across three models: a monolithic model, a traditional federated learning (FL) model, and the proposed MCFL model. Our results demonstrate that the MCFL model improves the prediction accuracy of the standard FL model by 13.86%, suggesting it as a promising enhancement in federated settings. Furthermore, MCFL is particularly well-suited for district heating forecasting, as it handles data heterogeneity, reinforces privacy protections, and supports scalability, making it an ideal choice for complex, distributed environments. • Multi-center Federated Learning (MCFL) applied to District Heating (DH) forecast. • MCFL balances privacy and accuracy in DH demand forecasting. • MCFL outperforms Federated Learning (FL) while maintaining its advantages. • The MCFL model offers scalability for various energy forecasting scenarios.