Buildings stop merging forecasts blindly: a local cross-fit gate avoids the 3% negative transfer that unconditional federation caused on Electricity
Synopsis
Distributed building energy management depends on forecasts, yet federated forecasting is usually scored on error without testing the decisions those forecasts support. UtilityGossip keeps a private and a collaborative low-rank adapter at each node and accepts a peer adapter only when randomized four-fold cross-fitting on local held-out data shows a meaningful error reduction. On Electricity, unconditional federation raised mean squared error more than 3% above local training while the gated method avoided that degradation; on Solar-Energy at the 96-step horizon it lowered mean squared error by 1.59% relative to local training using half the communication of federated averaging.
Interpretation
Forecast-model sharing between buildings now lets the receiver decide whether to adopt an update instead of merging whatever arrives. FedAvg and FedGossip previously merged neighbor or global adapters unconditionally, which can hurt an individual site when building load distributions differ. On the Electricity benchmark, FedAvg and FedGossip raised mean squared error by 3.46% and 3.23% relative to local training, while UtilityGossip avoided that degradation and produced the lowest federated mean squared error and the lowest worst-node mean squared error.
Each node keeps two rank-8 low-rank adapters around one frozen PatchTST backbone, and only the collaborative adapter is exchanged while the private adapter never leaves the node. Low-rank adaptation represents a dense prediction-head update with two thin matrices, so the H=96, d=5632, r=8 configuration learns 45,824 values instead of the 540,672 a dense update requires. The receiver runs randomized four-fold cross-fitting on 128 locally held-out rows and accepts a candidate only when the mean gain to uncertainty ratio exceeds 2.0, the standardized effect exceeds 0.3 and a relative-gain floor is met, with evaluation on a final test segment the gate cannot access.
Better forecast accuracy did not translate into proportional electricity savings, so prediction error and operating value are reported separately. Replaying every forecast source through the same frozen battery policy separates the contribution of battery flexibility, learned forecasting and federation. In the held-out week replay on Electricity, the UtilityGossip forecast supported a 19.34% cost reduction, 3.24% lower grid imports and 9.89% lower carbon emissions relative to no storage, but the additional control gain over a strong local forecast was not statistically resolved across three seeds (p=0.275).
Perspective
The demonstrated scope is the public Electricity and Solar-Energy benchmarks, with nodes formed as 32 client portfolios by clustering fitting-period spectral fingerprints rather than individual meters; battery control is replayed only on Electricity using one frozen PPO policy trained on a 672-hour calibration segment and evaluated on the following 168-hour held-out week. The district-peak guard acts only on aggregate active power and does not represent voltage, phase imbalance, reverse flow, feeder thermal limits or protection operation.
The gate is a model-selection rule rather than a sequential hypothesis test, applied adaptively over 100 rounds without multiplicity correction, so acceptance does not guarantee lower error on unseen data. In the privacy sensitivity study the acceptance rate falls from 32.9% to about 20%, the strongest setting still has epsilon = 14.13, and the guarantee covers one node's transmitted stream without later peer re-sharing. With three paired seeds, normality cannot be assessed credibly and the paired p values are descriptive.