异构动态数据环境下分布式光伏在线增量概率功率预测

Abstract Data sharing is a standard solution to improve the prediction accuracy of data-driven models for distributed photovoltaic (PV) power with small samples. Unfortunately, in practice, due to decentralized ownership and diverse, dynamic external environments, this solution suffers from challenges in data privacy, heterogeneity, and dynamic data learning. To handle these challenges, this paper proposes an incremental probabilistic prediction method based on a Bayesian stochastic configuration network (BSCN) and personalized federated learning (PFL). Concretely, a stochastic configuration network, an emerging neural network with a single hidden layer and no iteration, is used to quickly build the power predictor. Aiming to obtain the posterior distribution and determine the probabilistic output, Bayesian inference is used to evaluate the output parameters of SCN. Faced with the performance degradation caused by small samples and heterogeneous data, a novel PFL framework is designed to improve the prediction accuracy while protecting privacy. Technically, the server acts as a bridge for information sharing and aggregates local posterior distributions in a personalized manner, guided by Wasserstein distance to integrate similar features as much as possible. With the personalized posterior from the server as the prior, each client performs personalized retraining locally to mitigate the adverse effects of the data heterogeneity while learning shared information from other clients. Moreover, an incremental learning strategy is proposed and seamlessly embedded into the PFL framework to continuously learn new modes without forgetting in dynamic environments. Extensive experiment results using public datasets demonstrate that the proposed method exhibits competitive probabilistic prediction performance compared to several state-of-the-art solutions for distributed PVs in the presence of small-sample, heterogeneous, and dynamic data.