FedBuild: Privacy-Preserving Federated Learning for Building Energy Forecasting

FedBuild implements a production-grade federated learning system with formal differential privacy (DP) guarantees. The system trains CNN-LSTM models across 50 buildings using the Flower framework, applies per-sample gradient clipping and Gaussian noise via Opacus, and evaluates three aggregation strategies (FedAvg, FedBN, FedProx) across privacy budgets from ε=0.5 to ε=6.5.

Key finding: DP-SGD gradient clipping acts as implicit regularization in federated settings with inter-client heterogeneity, causing all DP configurations to outperform the non-private federated baseline—a novel effect unreported in prior building-energy forecasting literature.

Features

• Flower-based orchestration: 50 clients, 10 sampled per round, 50 federation rounds
• DP-SGD via Opacus: Per-sample gradient clipping (C=1.0) + Gaussian noise, server-side RDP composition accounting
• Three strategies evaluated: FedAvg, FedBN (local batch norm), FedProx (proximal regularization)
• 12 privacy configurations: 3 strategies × 4 ε targets (0.5, 1.0, 3.0, 6.5)
• DP-compatible architecture: CNN-LSTM with DPLSTM + GroupNorm substitutions
• Real-world dataset: Building Data Genome Project 2, 50 office/education buildings (Panther site)
• Reproducibility: Round-by-round checkpointing, seed fixing, aggregated metrics + per-run histories
• Full privacy validation: Server-side RDP composition with explicit final ε reporting (achieved ≈ target within 3%)
• Publication-ready outputs: 9 figures (PDF+PNG), methodology document, comprehensive tables