Conference

Abstract

Accurate peak load forecasting is critical to power system operation, but household-level forecasting remains difficult due to load variability. Since consumption data are privacy-sensitive, federated learning (FL) offers a natural framework for collaborative modeling without centralized data collection. We propose a peak-aware FL framework with two components: (i) a peak-weighted loss that emphasizes peak-time samples during training, and (ii) a dual-path architecture that combines a shared vector-quantized (VQ) codebook for federated pattern exchange with a lightweight DLinear residual for per-household personalization. We evaluate models using PAPE (Peak Absolute Percentage Error) and HR (Hit Rate), metrics tailored to peak-region accuracy. Across a broad set of baselines, the proposed model achieves the best peak accuracy, reducing PAPE by 9.8% relative to local-only training. Notably, our light-weight model surpasses time-series foundation models, showing that loss design and personalization outweigh model scale for household peak forecasting under privacy constraints.