Distributed photovoltaic power forecasting based on personalized federated adversarial learning

Abstract

Existing distributed photovoltaic (PV) power forecasting methods fail to address the impact of sample scarcity and heterogeneity in PV power data. Moreover, training a single prediction model proves challenging to meet the personalized forecasting needs of different PV stations in distributed environments. This paper proposes a personalized federated generative adversarial network (PFedGAN)-based DPV power forecasting method. Leveraging the federated learning (FL) framework, it achieves collaborative training of prediction models among DPV stations while preserving data privacy. y introducing generative adversarial networks (GAN) and personalized strategy optimization into the FL training process, it alleviates data scarcity issues and reduces the impact of data heterogeneity. A TimesNet-DeepAR (TNE-DeepAR) power prediction model is designed, where the TimesNet module extracts correlations between PV power data from different time periods, and the DeepAR module facilitates PV power prediction, mitigating the effects of meteorological factors' multi-periodic variations on PV power. Experimental results show that the proposed hybrid prediction model reduces the average mean absolute percentage error (MAPE) by 30–40 % compared to single models. The proposed approach reduces the MAPE by 9.79 % compared to traditional methods and by 49.62 % for PV stations with scarce data.