Abstract
Electric load forecasting is essential for power management and stability insmart grids. This is mainly achieved via advanced metering infrastructure,where smart meters (SMs) record household energy data. Traditional machinelearning (ML) methods are often employed for load forecasting, but require datasharing, which raises data privacy concerns. Federated learning (FL) canaddress this issue by running distributed ML models at local SMs without dataexchange. However, current FL-based approaches struggle to achieve efficientload forecasting due to imbalanced data distribution across heterogeneous SMs.This paper presents a novel personalized federated learning (PFL) method forhigh-quality load forecasting in metering networks. A meta-learning-basedstrategy is developed to address data heterogeneity at local SMs in thecollaborative training of local load forecasting models. Moreover, to minimizethe load forecasting delays in our PFL model, we study a new latencyoptimization problem based on optimal resource allocation at SMs. A theoreticalconvergence analysis is also conducted to provide insights into FL design forfederated load forecasting. Extensive simulations from real-world datasets showthat our method outperforms existing approaches regarding better loadforecasting and reduced operational latency costs.