A short-term wind power forecasting method based on multivariate signal decomposition and variable selection.

Accurate and effective short-term wind power forecasting is vital for the large-scale integration of wind power generation into the power grid. However, due to the intermittence and volatility of wind resources, short-term wind power forecasting is challenging. To address the issue that the existing decomposition forecasting methods ignore the coupling relationship between wind power series and multiple meteorological series, this study proposes a short-term wind power forecasting method based on multivariate signal decomposition and variable selection. First, multivariate variational mode decomposition (MVMD) is used to perform time-frequency synchronous analysis on wind power and multidimensional meteorological series, thereby decomposing them into the same predefined number of frequency-aligned intrinsic mode functions (IMFs). Secondly, elastic net (EN) is used for supervised variable selection on all IMFs to provide a high-quality training set for the forecasting model, thereby enhancing precision and interpretability. Next, a hybrid deep neural network combining convolutional neural network (CNN), bidirectional long-short term memory (BiLSTM) neural network, and multi-head attention (MHA) mechanism is employed to model the output curve of a group of wind turbines in a wind farm. Finally, the proposed method is comprehensively evaluated through four sets of comparative experiments and multiple evaluation metrics on data gathered from the Mahuangshan first wind farm in China with four forecasting horizons: 15-min ahead, 30-min ahead, 45-min ahead, and 1-h ahead. The experimental results show that the proposed method significantly outperforms fifteen existing deep learning methods in terms of precision and stability. • A novel multivariate signal decomposition strategy is proposed for wind power forecasting. • Elastic net is firstly applied to provide a high-quality training set for the forecasting model. • Utilizing the multi-head attention mechanism to capture the long-term dependencies in wind power sequence. [ABSTRACT FROM AUTHOR]