Point and interval forecasting of ultra-short-term wind power based on a data-driven method and hybrid deep learning model

Abstract

Accurate and reliable wind power forecasting (WPF) is significant for ensuring power systems’ economic operation and safe dispatching and for reducing the technical and economic risks faced by power market participants. Based on data-driven and deep-learning methods, we propose a hybrid ultra-short-term WPF framework that can achieve accurate point and interval WPF. First, the multi-sourced and multi-dimensional data sets of wind power plant are preprocessed. Second, feature selection (FS) is conducted to eliminate redundant features. Third, the wind power sequence is decomposed through the variational modal decomposition improved by grey wolf optimization (GWO-VMD). Then, the BiLSTM-Attention model is established to predict each subsequence of wind power. Finally, the prediction intervals of wind power under different confidence levels are estimated by kernel density estimation with the Gaussian kernel function (KDE-Gaussian). The proposed FS-GWO-VMD-BiLSTM-Attention forecasting framework is compared with benchmark models to verify its practicability and reliability. Compared with the BPNN, the mean absolute error, mean absolute percentage error, and mean square error of the FS-GWO-VMD-BiLSTM-Attention model are reduced by 94.03%, 85.82%, and 99.51%, respectively. Furthermore, according to the coverage width-based criterion, KDE-Gaussian is superior to other interval forecasting methods, which can achieve more reliable forecasting of prediction interval.