Privacy-preserving energy analytics in smart offices via container-based Federated Learning

Abstract

Federated Learning (FL) has emerged as a promising paradigm to enable privacy-preserving machine learning across distributed IoT devices. This work relies on SimulaFed, a container-based in-simulation framework for FL that is readily applicable to IoT scenarios. It leverages real-world energy data from an office building in which environmental and occupancy parameters were monitored by an IoT system. Our framework performs distributed model training that preserves occupant privacy without incurring prohibitive communication overhead and benchmarks four aggregation rules–Federated Averaging (FedAvg), Federated Proximal (FedProx), FedAdam, and SCAFFOLD.

Using $\approx$ 262 000 hourly windows and a lightweight 1-D CNN ($\approx$  0.35 M parameters; 354 488 weights), we benchmarked four aggregation rules. FedProx, with a tuned proximity term ($\mu =0.05$), achieved the lowest MAE: 0.755 ± 0.000, marginally ahead of FedAvg (0.764 ± 0.084) by 1.2%. SCAFFOLD delivered accuracy comparable to FedAvg (MAE $0.771\pm 0.042$) but with a higher runtime footprint; FedAdam increased computational cost without accuracy gains. Each update payload is about 1.4 MB per client; across 17 clients and 10 rounds (upload + broadcast) this totals $\approx$480 MB. Detailed CPU/memory telemetry is reported in Section 4 and Table 13.

These results confirm the viability of SimulaFed as a rapid-prototyping platform for energy-aware FL in smart offices, paving the way for deployments that balance data confidentiality, prediction accuracy and resource usage.