Short-term Load Forecasting of Distribution Transformer Supply Zones Based on Federated Model-Agnostic Meta Learning

Abstract

With the increasing data privacy concerns raised by not only organizations but also individuals in distribution systems, traditional centralized data-driven forecasting approaches for short-term load forecasting (STLF) in distribution transformer supply zones are confronted with the predicament of isolated data island. To this end, a federated model-agnostic meta learning (FMAML) based STLF method is proposed. On the basis of federated learning(FL), model agnostic meta learning (MAML) is employed to build high-quality personalized models for clients, thereby significantly enhancing the personalization and compatibility of the Federated Learning FL, while easing data privacy concerns leveraging the feature of FL. The stochastic controlled averaging (SCA) algorithm is integrated as the federated aggregation algorithm to mitigate the impacts of client-drift (CD) phenomenon that causes slow convergence and even divergence during the training process, especially when the data is highly heterogeneous. Finally, numerical results verify the high accuracy and strong robustness to data heterogeneity and packet dropout of the proposed method.