Federated Deep Learning for Scalable and Explainable Load Forecasting in Privacy-Conscious Smart Cities
As smart cities evolve, energy infrastructures are becoming more decentralized and dynamic due to the increased integration of renewables, electric vehicles, and consumer-driven usage behaviors. These shifts introduce significant complexity into electricity load forecasting, challenging conventional...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11075650/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | As smart cities evolve, energy infrastructures are becoming more decentralized and dynamic due to the increased integration of renewables, electric vehicles, and consumer-driven usage behaviors. These shifts introduce significant complexity into electricity load forecasting, challenging conventional centralized AI models. Such models often face limitations in scalability, generalization across heterogeneous environments, and preserving user data privacy. Furthermore, their lack of interpretability impedes transparent decision-making, an essential requirement in critical energy management and policy planning contexts. To address these issues, this study proposes HHCTE-FL, a Hierarchical Hybrid Convolutional Transformer Extractor embedded within a federated learning framework. The architecture supports multi-horizon electricity load forecasting while ensuring model transparency through Layer-wise Interpretative Attention Maps (LIAM) and robustness via personalized federated optimization. Functional modules such as Dynamic Temporal Refinement (DTR) and Multi-Stage Adaptive Feature Selection (MSAFS) enable adaptive temporal modeling and feature prioritization across non-IID client data. Evaluation on real-world smart grid datasets demonstrates that HHCTE-FL consistently surpasses 12 benchmark methods, attaining 98.7% accuracy, with the lowest forecast errors (MAE: 0.081, RMSE: 0.114). The model achieves convergence within just 28 communication rounds and operates with low overhead, while attaining a Federated Stability Index (FSI) of 0.96, indicating high training consistency. Statistical significance analysis further affirms its superiority in performance and reliability. By integrating explainable deep learning with privacy-preserving federated learning, HHCTE-FL establishes a scalable and trustworthy paradigm for intelligent electricity load forecasting—aligned with the demands of modern urban energy systems and smart city sustainability goals. |
|---|---|
| ISSN: | 2169-3536 |