Digital twin-driven short-term photovoltaic power forecasting using a hybrid TCN-transformer with residual correction

Abstract

To address the significant impact of meteorological factors on photovoltaic (PV) power generation and the complex time-series characteristics, this paper proposes a short-term power prediction method incorporating digital twin concepts. First, a virtual-real mapping framework is constructed between the physical system and the virtual model to achieve dynamic perception and data-driven modeling of multi-source meteorological data. Based on this, a hybrid prediction model combining a Temporal Convolutional Network (TCN) and a Transformer is designed to enhance the modeling ability for local variations and global dependencies. Furthermore, a residual compensation mechanism based on similar meteorological scenarios is introduced to adaptively correct the initial prediction results. Experimental validation on a real PV power plant dataset demonstrates that this method exhibits good prediction performance and stability under various weather conditions, providing effective support for the operation optimization and scheduling decisions of PV power generation systems.