Private federated forecasting

Electricity forecasting is moving onto the meter itself, with researchers testing a setup that trains shared models without sending household usage logs to a central server. (arxiv.org) In this kind of system, each smart meter or local device trains on its own history and sends back model updates instead of raw readings. Federated learning was built for that arrangement, but load forecasting has been a hard case because homes do not use power in the same way. (arxiv.org) That mismatch is called heterogeneity: one home may peak at 6 p.m., another after midnight, and a third may have solar panels changing the pattern entirely. Researchers have been adding “personalization layers,” a split model that keeps some parameters local while sharing the rest, to stop one average model from fitting nobody well. (arxiv.org) A 2024 load-forecasting paper from the University of Alabama in Huntsville and Pacific Northwest National Laboratory used that personalized approach on smart-meter data and reported better accuracy than standard federated and local baselines under non-identical data conditions. The same paper frames smart-meter privacy as a practical issue, citing the 2009 Dutch court ruling that halted a mandatory smart-meter rollout over privacy concerns. (arxiv.org) The newer variant being discussed, DP-SA-FedPer, adds two more privacy tools on top of personalization: differential privacy and secure aggregation. Differential privacy mixes statistical noise into updates so one household is harder to single out, while secure aggregation hides each client’s update and reveals only the combined total to the server. (arxiv.org) That combination follows a broader pattern in power and machine-learning research. A 2022 University of Luxembourg study on residential short-term load forecasting found that federated learning paired with differential privacy and secure aggregation could preserve high forecasting performance while protecting both load data and model information. (arxiv.org) The appeal for utilities is straightforward: short-term load forecasts help plan, operate, and schedule power systems, especially as intermittent renewable generation makes demand balancing harder. Smart meters provide the detailed household data needed for those forecasts, but centralized collection creates ownership, privacy, and security risks. (arxiv.org) The open question is what privacy means once gradients, not raw data, become the thing being shared. A 2025 USENIX security paper says private training data can still be reconstructed from shared gradients in federated learning through gradient inversion attacks, and newer work has shown that even aggregated gradients can leak labels under some threat models. (usenix.org, arxiv.org) That is why newer papers are also focusing on verifiable secure aggregation, not just hidden aggregation. Recent work in IEEE and Springer venues describes schemes that let clients check whether a server computed the aggregate correctly without exposing individual updates, a response to the trust problem in server-mediated training. (ieeexplore.ieee.org, link.springer.com) So the story here is less about one accuracy number than about where forecasting happens. If the model can stay accurate while the data stays on the device, utilities get a usable forecast and households give up less of the record of how they live. (arxiv.org, arxiv.org)