Short-term Wind Power Prediction Based on DLinear-CrossFormer Considering Error Correction

Predicting wind power accurately is a vital method for mitigating challenges in integrating large-scale wind power into the grid, leading to enhanced security and cost-effectiveness in power system functionality. A new hybrid wind power forecasting framework is proposed to enhance accuracy of this paper's short-term wind power projection. Firstly, the original wind power series undergoes decomposition into trend components and remaining detail components by singular spectrum analysis (SSA). Considering the model characteristics, the DLinear is employed to forecast the series trend, and the CrossFormer predicts the remaining detail components containing nonlinearities and periodicities, and the initial prediction value is obtained by linear superposition. The Light Gradient Boosting Machine (LightGBM) model predicts the error series to correct and compensate the initial prediction value. To validate the predictive capabilities of the model, experiments were carried out using wind power data from four stations exhibiting high correlation. The findings demonstrate that the hybrid model outperforms a single deep learning model in terms of accuracy and stability when compared to existing single prediction models, which also suggest that the hybrid prediction framework is a robust and dependable approach for wind power prediction.