A dual-layer feature-selection transformer network for transferable probabilistic forecasting of PV power

Accurate and reliable forecasting lays a solid foundation for enhancing Photovoltaic (PV) integration and facilitating its participation in demand response programs. The influencing factors in the PV power forecasting contain multiple types of uncertainties that cannot be ignored in the pursuit of reliable forecasting results. Traditional deterministic forecasting methods are difficult to provide sufficient uncertainty information to assist decision making. In this paper, a novel Dual-Layer Feature-Selection Transformer Network (DLFS-TN) is proposed for probabilistic forecasting of PV power. The proposed model introduces a Dual-Layer Feature-Selection Machine (DLFSM) by combining higher-order partial correlation coefficients with neural networks to dynamically assign feature weights. Compared to the method without DLFSM, the prediction accuracy of the DLFS-TN is improved by 7.47 %, which is 4.87 % higher than the average prediction accuracy of the six benchmark models. Moreover, the case study demonstrates the strong generalization capability of DLFS-TN across PV stations in diverse regions, as well as its effective scenario transferability to wind power, electric vehicles, air conditioning, and integrated energy systems, achieving an average accuracy of 88.01 % across these four scenarios.