Federated Few-shot Learning for Short-term Energy Information Forecasting in the Electric IoT

Federated learning (FL) has found substantial application within the Electric Internet of Things (EIoT). Challenges arise from intermittent data collection and unstable energy sources, leading to a prevalence of non-independent and identically distributed (non-IID) clients with sparse data. This paper proposes a novel communication-friendly FL approach for few-shot short-term forecasting in the EIoT that leverages knowledge distillation from distributed local models and employs adversarial training to produce synthetic samples, thereby facilitating client model training through data augmentation. The proposed algorithm demonstrates considerable resilience to client data variability and its efficacy is corroborated with a real-world open-source load dataset.