New Initialization Strategies Boost Federated Load Forecasting Accuracy

Accurate short-term load forecasting (STLF) is vital for power systems but often requires sensitive smart-meter data, raising privacy concerns. Federated Learning (FL) offers a privacy-preserving solution, yet it struggles with the inherent heterogeneity of client load data, where different users respond uniquely to external factors and have distinct temporal profiles. This heterogeneity can significantly degrade forecasting performance. To address this, the research highlights the critical role of model initialization in federated STLF and introduces two new strategies. The first is a "pretrained initialization" for the global model, leveraging auxiliary public data to reduce client drift during training. The second, SLIAvg, is a sequential local initialization strategy that allows clients to start each communication round with progressively adapted models, promoting more consistent training. These initialization methods are compatible with existing FL frameworks and privacy techniques, demonstrating improved forecasting performance, reduced client drift, and better convergence on real smart-meter data across two forecasting architectures.