A novel multi-step ahead solar power prediction scheme by deep learning on transformer structure

Photovoltaic (PV) power generation inherently possesses uncertainty and is susceptible to significant short-term fluctuations, posing a notable risk to power grid stability. To address this challenge, accurate solar irradiance prediction emerges as a viable solution to mitigate power intermittency. In particular, the complexity increases when considering multistep prediction as opposed to single-step prediction. Consequently, the pursuit of effective multi-step prediction methods becomes a pressing and essential research endeavor. This paper introduces a novel approach for multi-step solar prediction (MSSP) model, founded upon the transformer framework. This model adeptly captures prolonged dependencies within solar data, thus accommodating trend variations. The MSSP model innovatively integrates a distilling operation and a generative decoder, which serve to reduce error propagation, construct replicas, and enhance model generalization and robustness. The experimental results show that the MSSP prediction range has minimal error accumulation from the first step to the tenth step the MAE and the MSE increase by only 0.3% and -6%. In the tenth step prediction, the MAE and MAPE are improved by 55.4% and 28.9% compared to the LSTM and the BiLSTM. The case study in the electricity market indicates that the MSSP reduces the costs of PV generators by 37.54% compared to the original method; The proposed model has highly prediction accuracy and powerful practicability, easy to be applied in practical engineering applications.