A hybrid attention-based deep learning approach for wind power prediction

Renewable energy, especially wind power, is a practicable and promising solution to mitigate the existing dilemma associated with climate change. Efficient and accurate prediction of wind power could guide a variety of decisions for resource management. To improve the accuracy of wind power prediction, most existing studies are multistage, where signal processing methods are first employed to decompose a single time series, and then deep learning methods are utilized for prediction. The aforementioned approaches have shown satisfactory results but tend to involve a burdensome time series decomposition process. To address this problem, this paper proposes a hybrid attention-based deep learning approach to achieve more efficient and accurate wind power prediction. The essential idea behind the proposed approach is to incorporate the cumbersome decomposition process into a hybrid deep learning model consisting of different deep neural networks, where each deep neural network is designed to perform a specific part of the prediction task to maximize its corresponding advantages. Compared with the typical deep learning models for time series prediction, e.g., long short-term memory (LSTM) and gated recurrent units (GRU), the proposed deep learning model has the following two major advantages: (1) the model eliminates the time series decomposition process by time embedding layers to achieve efficient prediction, and (2) the model achieves more powerful high-level temporal feature extraction by leveraging the combination of a convolutional neural network (CNN), multiple stacked bidirectional long short-term memory (Bi-LSTM) networks, and the attention mechanism, thus providing high accuracy prediction. The proposed method is evaluated with a real-world wind power dataset in Turkey, and comparative experiments demonstrate the effectiveness and applicability of the proposed method. • A hybrid attention-based deep learning model was developed for wind power prediction. • The time embedding layer improves the accuracy and efficiency of the prediction. • The combination of two deep neural networks improves the prediction accuracy. • An attention mechanism reduces the loss of information. • The predicted results are well consistent with the observed dataset.