Centralized sequential federated learning: single-server simulation for cross-region load forecasting

Document Type

Article

Publication Date

4-1-2026

Abstract

To address data distribution heterogeneity (non-IID) and training-efficiency challenges in cross-regional electric load forecasting, this paper proposes a Centralized Sequential Federated Learning (CSFL) framework. In a single-server, centrally stored setting, CSFL emulates multi-regional “clients” via logical partitions, employs sequential local training with central aggregation, and incorporates a dynamic learning-rate decay coupled with round resets to promote progressive integration of cross-regional features. Combined with a feature-engineering pipeline based on Ward's minimum-variance clustering and Pearson correlation analysis, CSFL substantially improves cross-regional forecasting performance. Experiments show that, compared with directly applying a single local model across regions, CSFL reduces the average forecasting error by 22.3%, and its gains are statistically significant according to five independent runs with a paired t-test (p=0.032). The method achieves efficient cross-regional knowledge fusion in a single-server environment, offering a high-performing and easily deployable solution for power grid dispatch centers.

Keywords

Centralized sequential federated learning (CSFL), Cross-regional forecasting, Electric load forecasting, Non-IID data

Publication Title

Pervasive and Mobile Computing

ISSN

1574-1192

DOI

10.1016/j.pmcj.2026.102176

Volume

117

Publisher

Elsevier B.V.

Share

COinS