Short-term photovoltaic power prediction model based on feature construction and improved transformer

The accurate predictions of photovoltaic (PV) power generation are crucial for the efficient operation, management, and dispatch of power systems. This study presents a short-term PV hybrid forecasting model that leverages feature construction and an improved Transformer architecture. The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm is optimized using the FOX optimization method to obtain smooth intrinsic modal function (IMF), which are combined with meteorological data to extract key features. Furthermore, a PV power prediction model is constructed based on an iTransformer network, and an attention mechanism is incorporated to capture frequency information between multivariate features, thereby enhancing the capability of prediction. To validate the effectiveness of the proposed method, experiments were conducted using actual PV data from the desert knowledge Australia solar centre (DKASC). The results indicate that the prediction performance of the proposed method is improved, with a root mean square error (RMSE) of 0.449 kW and a mean absolute error (MAE) of 0.333 kW. Compared to the iTransformer model, the proposed method achieves improvements in RMSE and MAE by 25.6 % and 17.7 %, respectively. The application of the proposed model to different PV generation data from DKASC demonstrates its accuracy, applicability, and usefulness.