Utility-Gated Personalized Federated Forecasting for Peak-Constrained Building Energy Management

Abstract Distributed building energy management depends on forecasts, yet federated forecasting is commonly evaluated without testing the decisions that those forecasts support. This study examines whether selective, communication-efficient collaboration improves both load prediction and battery operation. The proposed UtilityGossipmethod maintains separate private and collaborative low-rank adapters at each node. A peer adapter is incorporated only when randomized cross-fitting on local held-out data indicates a meaningful reduction in forecast error. Evaluation on public electricity and solar benchmarks covers longer horizons, alternative network topologies, node failures, and transmit-time noise. At the 96-step horizon, the method reduces Solar-Energy mean squared error by 1.59% relative to local training and avoids the greater than 3% degradation caused by unconditional federation on Electricity, using half the communication of federated averaging. Operational value is assessed by replaying every forecast source through the same frozen battery policy for a held-out week. An external projection prevents coordinated charging from increasing the running no-storage district peak. The proposed forecast supports a 19.34% cost reduction relative to no storage, and diagnostic replays confirm that forecast timing affects dispatch. Its additional control gain over a strong local forecast is not statistically resolved. The results support selective federation for distributed forecasting but show that lower prediction error does not necessarily yield proportional energy savings.