Hierarchical model based on optimized feature decomposition and deep learning for short-term wind-power forecasting

Abstract

Renewable energy development relies heavily on accurate wind-power forecasting. However, predicting wind power presents significant challenges, given the unique operational complexity inherent in wind farms. To overcome these challenges, this study proposes a novel hierarchical model based on optimized feature decomposition and deep learning. First, variational mode decomposition (VMD) is performed to decompose wind energy to mitigate variability and instability. Then, the Rime optimization algorithm (RIME) is implemented to optimize the parameters of VMD, thereby enhancing the effective decomposition of wind power into multiple, smoothly varying modal components. These components and the selected meteorological features are then used to generate sequential data, which are input into a temporal convolutional network (TCN) to extract time-series information from the wind-power data. A bidirectional long short-term memory network (BiLSTM) with self-attention mechanism (Attention) is incorporated to capture both long-term and more complex temporal patterns. During the model-training phase, predictions from the validation set are used to optimize the TCN hyperparameters via the RIME algorithm. Finally, the optimized model is tested on a dataset of forecast wind power. The results show that, compared to the TCN–BiLSTM–Attention model, the root mean square error and mean absolute error of the proposed method are lower by 54.54% and 50.6%, respectively, which verifies the superior prediction accuracy of the proposed model.