Federated meta-learning with gradient-parameter aggregation for load forecasting under data scarcity

Short-term load forecasting is critical for grid stability and energy management, yet growing privacy concerns and data ownership regulations increasingly restrict the centralized data aggregation that current methods rely on. Federated learning enables model training without raw data sharing, but existing federated forecasting methods assume sufficient local data at each client, an assumption that does not hold for all individual residential consumers. To address the need for distributed learning with limited data, this paper presents FedGPA ( Fed erated Meta-Learning via G radient- P arameter A ggregation), which trains a single global model under federated few-shot conditions using both gradient-based and parameter-based knowledge. FedGPA introduces a two-stage knowledge aggregation procedure, Multi-Perspective Weighted Aggregation, and Gradient-Parameter Fusion to balance fast adaptation with global stability across distributed, data-scarce nodes. Specifically, MPWA balances uniform, loss-weighted, and distance-weighted aggregation signals to improve global model consistency, while GPF adaptively combines gradient and parameter knowledge to enable stable convergence under data scarcity. Experiments on real-world residential households compare FedGPA against centralized and federated baselines using LSTM and Transformer architectures across varying forecasting horizons. FedGPA consistently outperforms all baselines across configurations while preserving data privacy.