A Transferable Framework of PV Power Forecasting for Cross-Regional Distributed PV Systems Using Domain Adversarial Temporal Network

The lack of meteorological forecast data has increased the inaccuracy of output power forecasting in distributed photovoltaic systems. Especially, for newly built distributed sites across regions, modeling based on data-driven methods is limited by insufficient historical data. Therefore, a domain adversarial temporal network (DATN) based transfer learning (TL) framework is proposed, which contains two main modules, power temporal forecaster and domain classifier. Among them, the domain classifier considering the hidden layer weights of long short-term memory network is designed to reduce the distribution mismatch between source and target domains. The DATN employs a TL strategy of cross-domain adversarial pretraining with target-specific prediction tuning. In four cross-regional transfer experiments, the effects of domain adaptation methods and transfer strategies are compared. The breakthrough is that the transfer effect on different target data volumes is analyzed for the first time. The results prove that the proposed transferable framework DATN consistently performs best.

© 2025 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission.