Privacy-Preserving Federated Learning for Building Energy Forecasting: A Differential Privacy Analysis of Aggregation Strategies

Forecasting hourly electricity consumption across distributed building portfolios is critical for grid management and demand response, yet the sensitivity of granular energy data constrains centralised aggregation. Federated learning (FL) trains models in situ, but the cost of layering formal differential privacy (DP) guarantees atop FL remains underexplored for building-energy workloads. We train a CNN-LSTM forecasting model under three FL aggregation strategies (FedAvg, FedProx, FedBN) across four cumulative privacy budgets (ε ∈ {0.5, 1, 3, 6.5}, δ = 10−5) over a federation of 50 commercial buildings from the Building Data Genome Project 2. Counter-intuitively, all twelve DP-FL configurations achieve lower validation mean squared error (MSE) than the FedAvg-without-DP baseline, by 17% to 37%. We advance per-sample gradient clipping in differentially private stochastic gradient descent (DP-SGD), which constrains heterogeneous client updates, as the most plausible mechanism, while noting that the DP-compatible architecture it requires is an uncontrolled confound. Within the DP regime, FedAvg exhibits the expected monotonic privacy–utility frontier, FedBN displays weakly anti-monotonic behaviour, and FedProx (μ = 0.01) fails to converge at all but the tightest budget, so aggregation-strategy choice can rival the privacy budget in its effect on accuracy. These are single-seed results from one site, offered as a preliminary characterisation that motivates the multi-seed replication and ablation we specify.